GPT-5 Mini

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'openai:gpt@5-mini',
  settings: {
    systemPrompt: 'You are a senior museum interpretation writer. Turn curator-approved source material into historically responsible, accessible audio storytelling. Preserve factual distinctions, never invent quotations or personal details, and write for the ear rather than the page.',
    maxTokens: 6000,
    thinkingLevel: 'medium'
  },
  messages: [
    {
      role: 'user',
      content: 'Write the complete audio guide for “Hidden in Plain Sight: Women Codebreakers of World War II,” a temporary exhibition at the Mid-Atlantic Museum of History.\n\nAUDIENCE AND FORMAT\n- General visitors ages 14 and older; no prior knowledge assumed\n- Six tracks, each 90–120 seconds when read aloud\n- Approximately 1,050–1,300 words total\n- Warm, precise, quietly suspenseful tone\n- One narrator throughout\n- Write for natural speech using short sentences, varied rhythm, and occasional direct questions\n- Add restrained production cues in brackets only where sound materially improves comprehension\n- End each track with a brief transition directing visitors to the next numbered object or area\n\nCURATOR-APPROVED FACT PACKET\n1. Entrance timeline: After the United States entered World War II, the Army and Navy rapidly expanded signals-intelligence operations. More than 10,000 American women served in wartime codebreaking roles. Many were recruited from colleges, teaching jobs, and clerical work because agencies sought language ability, mathematical skill, pattern recognition, accuracy, and discretion.\n2. Recruitment letter, 1942: The displayed letter invites a mathematics instructor to Washington for unspecified government work. It does not mention codebreaking. Wartime secrecy meant recruits often learned the nature of the job only after arrival. They signed secrecy oaths and were instructed not to discuss their work, including with family.\n3. Intercept worksheet: Operators copied groups of letters or numbers from enemy radio traffic. Analysts compared repeated patterns, call signs, transmission schedules, and earlier messages. Decryption was usually collaborative and incremental, not a lone flash of inspiration. A transcription error could waste hours or conceal a useful pattern.\n4. Two-rotor cipher machine: The exhibition contains a training device used to demonstrate rotor-based substitution. Rotors changed the electrical path with each keystroke, so the same letter could encrypt differently at different moments. The device is not an enemy machine captured in combat and must not be described as one.\n5. Pacific operations map: Navy codebreaking units helped identify patterns in Japanese naval communications. Intelligence derived from multiple sources contributed to operational decisions in the Pacific. Avoid claiming that any single decrypted message, person, or unit independently determined the outcome of a battle.\n6. Demobilization photograph, 1946: Many women left government service after the war as military organizations contracted and social expectations shifted. Secrecy prevented them from explaining the importance of their experience to prospective employers. Some later returned to teaching, raised families, entered technical professions, or continued in intelligence work.\n7. Oral-history station: Decades later, declassification and oral-history projects made fuller public recognition possible. The listening station includes recollections recorded late in life. Memory can be vivid but imperfect, so oral histories should be presented as personal testimony rather than treated as complete institutional records.\n\nREQUIRED TRACKS\n1. Orientation: secrecy, scale, and the exhibition’s central question\n2. The recruitment letter: being hired without a clear job description\n3. The intercept worksheet: how teamwork and precision shaped the work\n4. The rotor demonstrator: explain the basic concept without technical overload\n5. The Pacific map: show the value of intelligence while avoiding a simplistic “they won the war” claim\n6. Demobilization and recognition: why these stories remained obscure and how historians recover them\n\nINTERPRETIVE REQUIREMENTS\n- Use “women codebreakers” as a broad descriptive phrase, but acknowledge that roles varied and not every worker personally broke a cipher\n- Distinguish interception, transcription, traffic analysis, and decryption in plain language\n- Do not invent a composite character, dialogue, diary entry, quotation, or specific individual\n- Do not overstate causation or imply that intelligence alone decided military outcomes\n- Briefly acknowledge that the work supported warfare and therefore carried human consequences, without becoming graphic\n- On first use, explain “signals intelligence” in everyday language\n- Include one recurring verbal motif tied to noticing patterns; keep it subtle and use it no more than three times\n- Close with a reflective invitation to listen to the oral histories\n\nDeliver only the finished script. Label each section “Track 1” through “Track 6,” followed by a short title. Include pronunciation guidance only if genuinely needed.'
    }
  ]
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "openai:gpt@5-mini",
            "settings": {
                "systemPrompt": "You are a senior museum interpretation writer. Turn curator-approved source material into historically responsible, accessible audio storytelling. Preserve factual distinctions, never invent quotations or personal details, and write for the ear rather than the page.",
                "maxTokens": 6000,
                "thinkingLevel": "medium"
            },
            "messages": [
                {
                    "role": "user",
                    "content": "Write the complete audio guide for “Hidden in Plain Sight: Women Codebreakers of World War II,” a temporary exhibition at the Mid-Atlantic Museum of History.\n\nAUDIENCE AND FORMAT\n- General visitors ages 14 and older; no prior knowledge assumed\n- Six tracks, each 90–120 seconds when read aloud\n- Approximately 1,050–1,300 words total\n- Warm, precise, quietly suspenseful tone\n- One narrator throughout\n- Write for natural speech using short sentences, varied rhythm, and occasional direct questions\n- Add restrained production cues in brackets only where sound materially improves comprehension\n- End each track with a brief transition directing visitors to the next numbered object or area\n\nCURATOR-APPROVED FACT PACKET\n1. Entrance timeline: After the United States entered World War II, the Army and Navy rapidly expanded signals-intelligence operations. More than 10,000 American women served in wartime codebreaking roles. Many were recruited from colleges, teaching jobs, and clerical work because agencies sought language ability, mathematical skill, pattern recognition, accuracy, and discretion.\n2. Recruitment letter, 1942: The displayed letter invites a mathematics instructor to Washington for unspecified government work. It does not mention codebreaking. Wartime secrecy meant recruits often learned the nature of the job only after arrival. They signed secrecy oaths and were instructed not to discuss their work, including with family.\n3. Intercept worksheet: Operators copied groups of letters or numbers from enemy radio traffic. Analysts compared repeated patterns, call signs, transmission schedules, and earlier messages. Decryption was usually collaborative and incremental, not a lone flash of inspiration. A transcription error could waste hours or conceal a useful pattern.\n4. Two-rotor cipher machine: The exhibition contains a training device used to demonstrate rotor-based substitution. Rotors changed the electrical path with each keystroke, so the same letter could encrypt differently at different moments. The device is not an enemy machine captured in combat and must not be described as one.\n5. Pacific operations map: Navy codebreaking units helped identify patterns in Japanese naval communications. Intelligence derived from multiple sources contributed to operational decisions in the Pacific. Avoid claiming that any single decrypted message, person, or unit independently determined the outcome of a battle.\n6. Demobilization photograph, 1946: Many women left government service after the war as military organizations contracted and social expectations shifted. Secrecy prevented them from explaining the importance of their experience to prospective employers. Some later returned to teaching, raised families, entered technical professions, or continued in intelligence work.\n7. Oral-history station: Decades later, declassification and oral-history projects made fuller public recognition possible. The listening station includes recollections recorded late in life. Memory can be vivid but imperfect, so oral histories should be presented as personal testimony rather than treated as complete institutional records.\n\nREQUIRED TRACKS\n1. Orientation: secrecy, scale, and the exhibition’s central question\n2. The recruitment letter: being hired without a clear job description\n3. The intercept worksheet: how teamwork and precision shaped the work\n4. The rotor demonstrator: explain the basic concept without technical overload\n5. The Pacific map: show the value of intelligence while avoiding a simplistic “they won the war” claim\n6. Demobilization and recognition: why these stories remained obscure and how historians recover them\n\nINTERPRETIVE REQUIREMENTS\n- Use “women codebreakers” as a broad descriptive phrase, but acknowledge that roles varied and not every worker personally broke a cipher\n- Distinguish interception, transcription, traffic analysis, and decryption in plain language\n- Do not invent a composite character, dialogue, diary entry, quotation, or specific individual\n- Do not overstate causation or imply that intelligence alone decided military outcomes\n- Briefly acknowledge that the work supported warfare and therefore carried human consequences, without becoming graphic\n- On first use, explain “signals intelligence” in everyday language\n- Include one recurring verbal motif tied to noticing patterns; keep it subtle and use it no more than three times\n- Close with a reflective invitation to listen to the oral histories\n\nDeliver only the finished script. Label each section “Track 1” through “Track 6,” followed by a short title. Include pronunciation guidance only if genuinely needed."
                }
            ]
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "textInference",
      "taskUUID": "086d97da-c862-4aea-8ff6-833605d43c08",
      "model": "openai:gpt@5-mini",
      "settings": {
        "systemPrompt": "You are a senior museum interpretation writer. Turn curator-approved source material into historically responsible, accessible audio storytelling. Preserve factual distinctions, never invent quotations or personal details, and write for the ear rather than the page.",
        "maxTokens": 6000,
        "thinkingLevel": "medium"
      },
      "messages": [
        {
          "role": "user",
          "content": "Write the complete audio guide for “Hidden in Plain Sight: Women Codebreakers of World War II,” a temporary exhibition at the Mid-Atlantic Museum of History.\n\nAUDIENCE AND FORMAT\n- General visitors ages 14 and older; no prior knowledge assumed\n- Six tracks, each 90–120 seconds when read aloud\n- Approximately 1,050–1,300 words total\n- Warm, precise, quietly suspenseful tone\n- One narrator throughout\n- Write for natural speech using short sentences, varied rhythm, and occasional direct questions\n- Add restrained production cues in brackets only where sound materially improves comprehension\n- End each track with a brief transition directing visitors to the next numbered object or area\n\nCURATOR-APPROVED FACT PACKET\n1. Entrance timeline: After the United States entered World War II, the Army and Navy rapidly expanded signals-intelligence operations. More than 10,000 American women served in wartime codebreaking roles. Many were recruited from colleges, teaching jobs, and clerical work because agencies sought language ability, mathematical skill, pattern recognition, accuracy, and discretion.\n2. Recruitment letter, 1942: The displayed letter invites a mathematics instructor to Washington for unspecified government work. It does not mention codebreaking. Wartime secrecy meant recruits often learned the nature of the job only after arrival. They signed secrecy oaths and were instructed not to discuss their work, including with family.\n3. Intercept worksheet: Operators copied groups of letters or numbers from enemy radio traffic. Analysts compared repeated patterns, call signs, transmission schedules, and earlier messages. Decryption was usually collaborative and incremental, not a lone flash of inspiration. A transcription error could waste hours or conceal a useful pattern.\n4. Two-rotor cipher machine: The exhibition contains a training device used to demonstrate rotor-based substitution. Rotors changed the electrical path with each keystroke, so the same letter could encrypt differently at different moments. The device is not an enemy machine captured in combat and must not be described as one.\n5. Pacific operations map: Navy codebreaking units helped identify patterns in Japanese naval communications. Intelligence derived from multiple sources contributed to operational decisions in the Pacific. Avoid claiming that any single decrypted message, person, or unit independently determined the outcome of a battle.\n6. Demobilization photograph, 1946: Many women left government service after the war as military organizations contracted and social expectations shifted. Secrecy prevented them from explaining the importance of their experience to prospective employers. Some later returned to teaching, raised families, entered technical professions, or continued in intelligence work.\n7. Oral-history station: Decades later, declassification and oral-history projects made fuller public recognition possible. The listening station includes recollections recorded late in life. Memory can be vivid but imperfect, so oral histories should be presented as personal testimony rather than treated as complete institutional records.\n\nREQUIRED TRACKS\n1. Orientation: secrecy, scale, and the exhibition’s central question\n2. The recruitment letter: being hired without a clear job description\n3. The intercept worksheet: how teamwork and precision shaped the work\n4. The rotor demonstrator: explain the basic concept without technical overload\n5. The Pacific map: show the value of intelligence while avoiding a simplistic “they won the war” claim\n6. Demobilization and recognition: why these stories remained obscure and how historians recover them\n\nINTERPRETIVE REQUIREMENTS\n- Use “women codebreakers” as a broad descriptive phrase, but acknowledge that roles varied and not every worker personally broke a cipher\n- Distinguish interception, transcription, traffic analysis, and decryption in plain language\n- Do not invent a composite character, dialogue, diary entry, quotation, or specific individual\n- Do not overstate causation or imply that intelligence alone decided military outcomes\n- Briefly acknowledge that the work supported warfare and therefore carried human consequences, without becoming graphic\n- On first use, explain “signals intelligence” in everyday language\n- Include one recurring verbal motif tied to noticing patterns; keep it subtle and use it no more than three times\n- Close with a reflective invitation to listen to the oral histories\n\nDeliver only the finished script. Label each section “Track 1” through “Track 6,” followed by a short title. Include pronunciation guidance only if genuinely needed."
        }
      ]
    }
  ]'

runware run openai:gpt@5-mini \
  settings.systemPrompt="You are a senior museum interpretation writer. Turn curator-approved source material into historically responsible, accessible audio storytelling. Preserve factual distinctions, never invent quotations or personal details, and write for the ear rather than the page." \
  settings.maxTokens=6000 \
  settings.thinkingLevel=medium \
  messages.0.role=user \
  messages.0.content="Write the complete audio guide for “Hidden in Plain Sight: Women Codebreakers of World War II,” a temporary exhibition at the Mid-Atlantic Museum of History.

AUDIENCE AND FORMAT
- General visitors ages 14 and older; no prior knowledge assumed
- Six tracks, each 90–120 seconds when read aloud
- Approximately 1,050–1,300 words total
- Warm, precise, quietly suspenseful tone
- One narrator throughout
- Write for natural speech using short sentences, varied rhythm, and occasional direct questions
- Add restrained production cues in brackets only where sound materially improves comprehension
- End each track with a brief transition directing visitors to the next numbered object or area

CURATOR-APPROVED FACT PACKET
1. Entrance timeline: After the United States entered World War II, the Army and Navy rapidly expanded signals-intelligence operations. More than 10,000 American women served in wartime codebreaking roles. Many were recruited from colleges, teaching jobs, and clerical work because agencies sought language ability, mathematical skill, pattern recognition, accuracy, and discretion.
2. Recruitment letter, 1942: The displayed letter invites a mathematics instructor to Washington for unspecified government work. It does not mention codebreaking. Wartime secrecy meant recruits often learned the nature of the job only after arrival. They signed secrecy oaths and were instructed not to discuss their work, including with family.
3. Intercept worksheet: Operators copied groups of letters or numbers from enemy radio traffic. Analysts compared repeated patterns, call signs, transmission schedules, and earlier messages. Decryption was usually collaborative and incremental, not a lone flash of inspiration. A transcription error could waste hours or conceal a useful pattern.
4. Two-rotor cipher machine: The exhibition contains a training device used to demonstrate rotor-based substitution. Rotors changed the electrical path with each keystroke, so the same letter could encrypt differently at different moments. The device is not an enemy machine captured in combat and must not be described as one.
5. Pacific operations map: Navy codebreaking units helped identify patterns in Japanese naval communications. Intelligence derived from multiple sources contributed to operational decisions in the Pacific. Avoid claiming that any single decrypted message, person, or unit independently determined the outcome of a battle.
6. Demobilization photograph, 1946: Many women left government service after the war as military organizations contracted and social expectations shifted. Secrecy prevented them from explaining the importance of their experience to prospective employers. Some later returned to teaching, raised families, entered technical professions, or continued in intelligence work.
7. Oral-history station: Decades later, declassification and oral-history projects made fuller public recognition possible. The listening station includes recollections recorded late in life. Memory can be vivid but imperfect, so oral histories should be presented as personal testimony rather than treated as complete institutional records.

REQUIRED TRACKS
1. Orientation: secrecy, scale, and the exhibition’s central question
2. The recruitment letter: being hired without a clear job description
3. The intercept worksheet: how teamwork and precision shaped the work
4. The rotor demonstrator: explain the basic concept without technical overload
5. The Pacific map: show the value of intelligence while avoiding a simplistic “they won the war” claim
6. Demobilization and recognition: why these stories remained obscure and how historians recover them

INTERPRETIVE REQUIREMENTS
- Use “women codebreakers” as a broad descriptive phrase, but acknowledge that roles varied and not every worker personally broke a cipher
- Distinguish interception, transcription, traffic analysis, and decryption in plain language
- Do not invent a composite character, dialogue, diary entry, quotation, or specific individual
- Do not overstate causation or imply that intelligence alone decided military outcomes
- Briefly acknowledge that the work supported warfare and therefore carried human consequences, without becoming graphic
- On first use, explain “signals intelligence” in everyday language
- Include one recurring verbal motif tied to noticing patterns; keep it subtle and use it no more than three times
- Close with a reflective invitation to listen to the oral histories

Deliver only the finished script. Label each section “Track 1” through “Track 6,” followed by a short title. Include pronunciation guidance only if genuinely needed."

{
  "taskType": "textInference",
  "taskUUID": "086d97da-c862-4aea-8ff6-833605d43c08",
  "model": "openai:gpt@5-mini",
  "settings": {
    "systemPrompt": "You are a senior museum interpretation writer. Turn curator-approved source material into historically responsible, accessible audio storytelling. Preserve factual distinctions, never invent quotations or personal details, and write for the ear rather than the page.",
    "maxTokens": 6000,
    "thinkingLevel": "medium"
  },
  "messages": [
    {
      "role": "user",
      "content": "Write the complete audio guide for “Hidden in Plain Sight: Women Codebreakers of World War II,” a temporary exhibition at the Mid-Atlantic Museum of History.\n\nAUDIENCE AND FORMAT\n- General visitors ages 14 and older; no prior knowledge assumed\n- Six tracks, each 90–120 seconds when read aloud\n- Approximately 1,050–1,300 words total\n- Warm, precise, quietly suspenseful tone\n- One narrator throughout\n- Write for natural speech using short sentences, varied rhythm, and occasional direct questions\n- Add restrained production cues in brackets only where sound materially improves comprehension\n- End each track with a brief transition directing visitors to the next numbered object or area\n\nCURATOR-APPROVED FACT PACKET\n1. Entrance timeline: After the United States entered World War II, the Army and Navy rapidly expanded signals-intelligence operations. More than 10,000 American women served in wartime codebreaking roles. Many were recruited from colleges, teaching jobs, and clerical work because agencies sought language ability, mathematical skill, pattern recognition, accuracy, and discretion.\n2. Recruitment letter, 1942: The displayed letter invites a mathematics instructor to Washington for unspecified government work. It does not mention codebreaking. Wartime secrecy meant recruits often learned the nature of the job only after arrival. They signed secrecy oaths and were instructed not to discuss their work, including with family.\n3. Intercept worksheet: Operators copied groups of letters or numbers from enemy radio traffic. Analysts compared repeated patterns, call signs, transmission schedules, and earlier messages. Decryption was usually collaborative and incremental, not a lone flash of inspiration. A transcription error could waste hours or conceal a useful pattern.\n4. Two-rotor cipher machine: The exhibition contains a training device used to demonstrate rotor-based substitution. Rotors changed the electrical path with each keystroke, so the same letter could encrypt differently at different moments. The device is not an enemy machine captured in combat and must not be described as one.\n5. Pacific operations map: Navy codebreaking units helped identify patterns in Japanese naval communications. Intelligence derived from multiple sources contributed to operational decisions in the Pacific. Avoid claiming that any single decrypted message, person, or unit independently determined the outcome of a battle.\n6. Demobilization photograph, 1946: Many women left government service after the war as military organizations contracted and social expectations shifted. Secrecy prevented them from explaining the importance of their experience to prospective employers. Some later returned to teaching, raised families, entered technical professions, or continued in intelligence work.\n7. Oral-history station: Decades later, declassification and oral-history projects made fuller public recognition possible. The listening station includes recollections recorded late in life. Memory can be vivid but imperfect, so oral histories should be presented as personal testimony rather than treated as complete institutional records.\n\nREQUIRED TRACKS\n1. Orientation: secrecy, scale, and the exhibition’s central question\n2. The recruitment letter: being hired without a clear job description\n3. The intercept worksheet: how teamwork and precision shaped the work\n4. The rotor demonstrator: explain the basic concept without technical overload\n5. The Pacific map: show the value of intelligence while avoiding a simplistic “they won the war” claim\n6. Demobilization and recognition: why these stories remained obscure and how historians recover them\n\nINTERPRETIVE REQUIREMENTS\n- Use “women codebreakers” as a broad descriptive phrase, but acknowledge that roles varied and not every worker personally broke a cipher\n- Distinguish interception, transcription, traffic analysis, and decryption in plain language\n- Do not invent a composite character, dialogue, diary entry, quotation, or specific individual\n- Do not overstate causation or imply that intelligence alone decided military outcomes\n- Briefly acknowledge that the work supported warfare and therefore carried human consequences, without becoming graphic\n- On first use, explain “signals intelligence” in everyday language\n- Include one recurring verbal motif tied to noticing patterns; keep it subtle and use it no more than three times\n- Close with a reflective invitation to listen to the oral histories\n\nDeliver only the finished script. Label each section “Track 1” through “Track 6,” followed by a short title. Include pronunciation guidance only if genuinely needed."
    }
  ]
}

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'openai:gpt@5-mini',
  settings: {
    systemPrompt: 'You are a senior retail merchandising auditor. Analyze only what is visibly supported by the supplied image and the user\'s stated planogram requirements. Clearly distinguish confirmed observations from anything uncertain, obscured, or unreadable. Do not invent product names, prices, counts, or compliance issues. Write polished operational reports that a field sales manager can act on immediately.',
    thinkingLevel: 'medium',
    maxTokens: 3000
  },
  inputs: {
    images: [
      'https://assets.runware.ai/assets/inputs/abf30303-2d63-4603-9d7c-d2dbfab0d1cd.jpg'
    ]
  },
  messages: [
    {
      role: 'user',
      content: 'I need a shelf-compliance audit for our sparkling water set. The approved planogram requires: two adjacent facings each of Lime, Grapefruit, and Black Cherry; placement on the middle two shelves; shelf-edge price labels aligned beneath each flavor; and one promotional tag reading “2 for $5.” Can you assess the attached store photo?'
    },
    {
      role: 'assistant',
      content: 'Yes. I’ll compare the visible shelf against those requirements, flag only image-supported issues, note anything that cannot be verified, and prioritize corrective actions.'
    },
    {
      role: 'user',
      content: 'Store: Brookfield Market #184, Denver. Visit date: July 29, 2026. Produce the finished audit in Markdown with these sections: Executive Summary, Requirement-by-Requirement Findings, Visible Stock and Placement Issues, Corrective Actions, and Verification Gaps. For each planogram requirement, assign Compliant, Non-compliant, or Not verifiable and cite the visible evidence. Keep the report under 600 words and make the corrective actions specific enough for a merchandiser to execute during this visit.'
    }
  ]
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "openai:gpt@5-mini",
            "settings": {
                "systemPrompt": "You are a senior retail merchandising auditor. Analyze only what is visibly supported by the supplied image and the user's stated planogram requirements. Clearly distinguish confirmed observations from anything uncertain, obscured, or unreadable. Do not invent product names, prices, counts, or compliance issues. Write polished operational reports that a field sales manager can act on immediately.",
                "thinkingLevel": "medium",
                "maxTokens": 3000
            },
            "inputs": {
                "images": [
                    "https://assets.runware.ai/assets/inputs/abf30303-2d63-4603-9d7c-d2dbfab0d1cd.jpg"
                ]
            },
            "messages": [
                {
                    "role": "user",
                    "content": "I need a shelf-compliance audit for our sparkling water set. The approved planogram requires: two adjacent facings each of Lime, Grapefruit, and Black Cherry; placement on the middle two shelves; shelf-edge price labels aligned beneath each flavor; and one promotional tag reading “2 for $5.” Can you assess the attached store photo?"
                },
                {
                    "role": "assistant",
                    "content": "Yes. I’ll compare the visible shelf against those requirements, flag only image-supported issues, note anything that cannot be verified, and prioritize corrective actions."
                },
                {
                    "role": "user",
                    "content": "Store: Brookfield Market #184, Denver. Visit date: July 29, 2026. Produce the finished audit in Markdown with these sections: Executive Summary, Requirement-by-Requirement Findings, Visible Stock and Placement Issues, Corrective Actions, and Verification Gaps. For each planogram requirement, assign Compliant, Non-compliant, or Not verifiable and cite the visible evidence. Keep the report under 600 words and make the corrective actions specific enough for a merchandiser to execute during this visit."
                }
            ]
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "textInference",
      "taskUUID": "899ae74e-4f36-45a4-99c9-28ef88ec0a41",
      "model": "openai:gpt@5-mini",
      "settings": {
        "systemPrompt": "You are a senior retail merchandising auditor. Analyze only what is visibly supported by the supplied image and the user's stated planogram requirements. Clearly distinguish confirmed observations from anything uncertain, obscured, or unreadable. Do not invent product names, prices, counts, or compliance issues. Write polished operational reports that a field sales manager can act on immediately.",
        "thinkingLevel": "medium",
        "maxTokens": 3000
      },
      "inputs": {
        "images": [
          "https://assets.runware.ai/assets/inputs/abf30303-2d63-4603-9d7c-d2dbfab0d1cd.jpg"
        ]
      },
      "messages": [
        {
          "role": "user",
          "content": "I need a shelf-compliance audit for our sparkling water set. The approved planogram requires: two adjacent facings each of Lime, Grapefruit, and Black Cherry; placement on the middle two shelves; shelf-edge price labels aligned beneath each flavor; and one promotional tag reading “2 for $5.” Can you assess the attached store photo?"
        },
        {
          "role": "assistant",
          "content": "Yes. I’ll compare the visible shelf against those requirements, flag only image-supported issues, note anything that cannot be verified, and prioritize corrective actions."
        },
        {
          "role": "user",
          "content": "Store: Brookfield Market #184, Denver. Visit date: July 29, 2026. Produce the finished audit in Markdown with these sections: Executive Summary, Requirement-by-Requirement Findings, Visible Stock and Placement Issues, Corrective Actions, and Verification Gaps. For each planogram requirement, assign Compliant, Non-compliant, or Not verifiable and cite the visible evidence. Keep the report under 600 words and make the corrective actions specific enough for a merchandiser to execute during this visit."
        }
      ]
    }
  ]'

runware run openai:gpt@5-mini \
  settings.systemPrompt="You are a senior retail merchandising auditor. Analyze only what is visibly supported by the supplied image and the user's stated planogram requirements. Clearly distinguish confirmed observations from anything uncertain, obscured, or unreadable. Do not invent product names, prices, counts, or compliance issues. Write polished operational reports that a field sales manager can act on immediately." \
  settings.thinkingLevel=medium \
  settings.maxTokens=3000 \
  inputs.images.0=https://assets.runware.ai/assets/inputs/abf30303-2d63-4603-9d7c-d2dbfab0d1cd.jpg \
  messages.0.role=user \
  messages.0.content="I need a shelf-compliance audit for our sparkling water set. The approved planogram requires: two adjacent facings each of Lime, Grapefruit, and Black Cherry; placement on the middle two shelves; shelf-edge price labels aligned beneath each flavor; and one promotional tag reading “2 for \$5.” Can you assess the attached store photo?" \
  messages.1.role=assistant \
  messages.1.content="Yes. I’ll compare the visible shelf against those requirements, flag only image-supported issues, note anything that cannot be verified, and prioritize corrective actions." \
  messages.2.role=user \
  messages.2.content="Store: Brookfield Market #184, Denver. Visit date: July 29, 2026. Produce the finished audit in Markdown with these sections: Executive Summary, Requirement-by-Requirement Findings, Visible Stock and Placement Issues, Corrective Actions, and Verification Gaps. For each planogram requirement, assign Compliant, Non-compliant, or Not verifiable and cite the visible evidence. Keep the report under 600 words and make the corrective actions specific enough for a merchandiser to execute during this visit."

{
  "taskType": "textInference",
  "taskUUID": "899ae74e-4f36-45a4-99c9-28ef88ec0a41",
  "model": "openai:gpt@5-mini",
  "settings": {
    "systemPrompt": "You are a senior retail merchandising auditor. Analyze only what is visibly supported by the supplied image and the user's stated planogram requirements. Clearly distinguish confirmed observations from anything uncertain, obscured, or unreadable. Do not invent product names, prices, counts, or compliance issues. Write polished operational reports that a field sales manager can act on immediately.",
    "thinkingLevel": "medium",
    "maxTokens": 3000
  },
  "inputs": {
    "images": [
      "https://assets.runware.ai/assets/inputs/abf30303-2d63-4603-9d7c-d2dbfab0d1cd.jpg"
    ]
  },
  "messages": [
    {
      "role": "user",
      "content": "I need a shelf-compliance audit for our sparkling water set. The approved planogram requires: two adjacent facings each of Lime, Grapefruit, and Black Cherry; placement on the middle two shelves; shelf-edge price labels aligned beneath each flavor; and one promotional tag reading “2 for $5.” Can you assess the attached store photo?"
    },
    {
      "role": "assistant",
      "content": "Yes. I’ll compare the visible shelf against those requirements, flag only image-supported issues, note anything that cannot be verified, and prioritize corrective actions."
    },
    {
      "role": "user",
      "content": "Store: Brookfield Market #184, Denver. Visit date: July 29, 2026. Produce the finished audit in Markdown with these sections: Executive Summary, Requirement-by-Requirement Findings, Visible Stock and Placement Issues, Corrective Actions, and Verification Gaps. For each planogram requirement, assign Compliant, Non-compliant, or Not verifiable and cite the visible evidence. Keep the report under 600 words and make the corrective actions specific enough for a merchandiser to execute during this visit."
    }
  ]
}

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'openai:gpt@5-mini',
  settings: {
    systemPrompt: 'You are a senior nonprofit grant writer. Produce persuasive, funder-ready prose grounded strictly in the supplied facts. Never invent statistics, partnerships, commitments, or evaluation results. Preserve requested headings and limits, connect activities to measurable outcomes, and use clear language rather than inflated claims.',
    maxTokens: 5000,
    thinkingLevel: 'medium'
  },
  messages: [
    {
      role: 'user',
      content: 'I need help drafting a grant request for Harbor Skills Alliance, a workforce nonprofit in Baltimore. We are seeking support to expand our certified nursing assistant training program. What information do you need?'
    },
    {
      role: 'assistant',
      content: 'Please provide the funding request and grant period; community need; target population; program activities and timeline; recruitment and employer partners; projected outputs and outcomes; evaluation methods; organizational qualifications; total project budget and use of grant funds; and the funder\'s priorities, required headings, and word limit. I will avoid unsupported claims and mark any essential missing information.'
    },
    {
      role: 'user',
      content: 'Draft a complete grant narrative of 1,200–1,400 words for the Northstar Community Foundation under these exact headings: Community Need; Program Design; Outcomes and Evaluation; Organizational Capacity; Use of Funds; Sustainability. End with a concise two-sentence closing case for support.\n\nFacts:\n- Applicant: Harbor Skills Alliance, a Baltimore 501(c)(3) founded in 2014.\n- Request: $250,000 for a 12-month expansion of Pathways to Care, our certified nursing assistant training and placement program.\n- People served: 120 Baltimore City residents, primarily unemployed or underemployed adults ages 20–45. Recruitment will focus on residents of ZIP codes 21215, 21217, and 21223. Participants must be legally eligible to work and able to meet Maryland CNA training requirements.\n- Need: In the last program year, 438 people applied for 72 available seats. Do not add external labor-market statistics; explain the need using only this demand and the barriers listed below.\n- Participant barriers identified during intake: training tuition, transportation, dependent-care costs, limited experience navigating healthcare hiring, and loss of income while attending class.\n- Program model: Four cohorts of 30 participants. Each cohort receives eight weeks of state-approved CNA instruction delivered by Baltimore Allied Health Institute, followed by four weeks of paid clinical placement and structured job-search support.\n- Support: Tuition and exam fees are fully covered. Participants receive a $150 weekly training stipend for 12 weeks, an unlimited local transit pass, dependent-care assistance of up to $800, loaner scrubs, and exam-preparation tutoring.\n- Career services: Individual career plan, résumé and interview coaching, two employer hiring events per cohort, application assistance, and 90 days of post-placement retention coaching.\n- Confirmed employer partners: Harborview Senior Living, Eastgate Rehabilitation Center, and St. Anne\'s Home Health. Describe them only as confirmed recruitment and interview partners; do not promise hires.\n- Targets: 120 enrolled; at least 102 complete training; at least 96 take the Maryland CNA exam; at least 86 earn certification; at least 78 obtain CNA employment within 90 days of certification; at least 66 of those placed remain employed for 90 days.\n- Evaluation: Attendance records, training completion records, exam registration and results, participant support logs, employer-verified placement, and 30-, 60-, and 90-day retention checks. Program director reviews a dashboard monthly; leadership reviews results quarterly. Participant surveys are administered at enrollment, completion, and 90 days after placement.\n- Capacity: 11 full-time staff. Pathways to Care launched in 2019 and has served 310 participants to date. Current team includes a program director, two career coaches, and a benefits navigator. Baltimore Allied Health Institute supplies licensed instructors and the approved curriculum. The organization has managed state and private grants up to $600,000.\n- Project budget: $612,000 total. Northstar funds will support $96,000 in participant stipends, $48,000 in tuition and exam fees, $36,000 in transportation and dependent-care assistance, $42,000 for one additional career coach, $18,000 for tutoring and training supplies, and $10,000 for data collection and evaluation. Explain that the remaining $362,000 will come from committed public workforce funds, unrestricted operating revenue, and pending corporate grants. Do not imply pending funds are secured.\n- Sustainability: We will use completion, certification, placement, and retention evidence to pursue renewed public workforce contracts and healthcare-employer contributions. We will also test a cost-sharing model with employer partners, but no partner has committed to it yet.\n- Funder priorities: economic mobility, equitable access to careers, measurable outcomes, and removal of practical barriers.\n\nWrite polished, specific prose suitable for direct submission. Keep the tone confident but evidence-based, distinguish outputs from outcomes, make the logic between funding and results explicit, and do not use bullet points or invent facts.'
    }
  ]
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "openai:gpt@5-mini",
            "settings": {
                "systemPrompt": "You are a senior nonprofit grant writer. Produce persuasive, funder-ready prose grounded strictly in the supplied facts. Never invent statistics, partnerships, commitments, or evaluation results. Preserve requested headings and limits, connect activities to measurable outcomes, and use clear language rather than inflated claims.",
                "maxTokens": 5000,
                "thinkingLevel": "medium"
            },
            "messages": [
                {
                    "role": "user",
                    "content": "I need help drafting a grant request for Harbor Skills Alliance, a workforce nonprofit in Baltimore. We are seeking support to expand our certified nursing assistant training program. What information do you need?"
                },
                {
                    "role": "assistant",
                    "content": "Please provide the funding request and grant period; community need; target population; program activities and timeline; recruitment and employer partners; projected outputs and outcomes; evaluation methods; organizational qualifications; total project budget and use of grant funds; and the funder's priorities, required headings, and word limit. I will avoid unsupported claims and mark any essential missing information."
                },
                {
                    "role": "user",
                    "content": "Draft a complete grant narrative of 1,200–1,400 words for the Northstar Community Foundation under these exact headings: Community Need; Program Design; Outcomes and Evaluation; Organizational Capacity; Use of Funds; Sustainability. End with a concise two-sentence closing case for support.\n\nFacts:\n- Applicant: Harbor Skills Alliance, a Baltimore 501(c)(3) founded in 2014.\n- Request: $250,000 for a 12-month expansion of Pathways to Care, our certified nursing assistant training and placement program.\n- People served: 120 Baltimore City residents, primarily unemployed or underemployed adults ages 20–45. Recruitment will focus on residents of ZIP codes 21215, 21217, and 21223. Participants must be legally eligible to work and able to meet Maryland CNA training requirements.\n- Need: In the last program year, 438 people applied for 72 available seats. Do not add external labor-market statistics; explain the need using only this demand and the barriers listed below.\n- Participant barriers identified during intake: training tuition, transportation, dependent-care costs, limited experience navigating healthcare hiring, and loss of income while attending class.\n- Program model: Four cohorts of 30 participants. Each cohort receives eight weeks of state-approved CNA instruction delivered by Baltimore Allied Health Institute, followed by four weeks of paid clinical placement and structured job-search support.\n- Support: Tuition and exam fees are fully covered. Participants receive a $150 weekly training stipend for 12 weeks, an unlimited local transit pass, dependent-care assistance of up to $800, loaner scrubs, and exam-preparation tutoring.\n- Career services: Individual career plan, résumé and interview coaching, two employer hiring events per cohort, application assistance, and 90 days of post-placement retention coaching.\n- Confirmed employer partners: Harborview Senior Living, Eastgate Rehabilitation Center, and St. Anne's Home Health. Describe them only as confirmed recruitment and interview partners; do not promise hires.\n- Targets: 120 enrolled; at least 102 complete training; at least 96 take the Maryland CNA exam; at least 86 earn certification; at least 78 obtain CNA employment within 90 days of certification; at least 66 of those placed remain employed for 90 days.\n- Evaluation: Attendance records, training completion records, exam registration and results, participant support logs, employer-verified placement, and 30-, 60-, and 90-day retention checks. Program director reviews a dashboard monthly; leadership reviews results quarterly. Participant surveys are administered at enrollment, completion, and 90 days after placement.\n- Capacity: 11 full-time staff. Pathways to Care launched in 2019 and has served 310 participants to date. Current team includes a program director, two career coaches, and a benefits navigator. Baltimore Allied Health Institute supplies licensed instructors and the approved curriculum. The organization has managed state and private grants up to $600,000.\n- Project budget: $612,000 total. Northstar funds will support $96,000 in participant stipends, $48,000 in tuition and exam fees, $36,000 in transportation and dependent-care assistance, $42,000 for one additional career coach, $18,000 for tutoring and training supplies, and $10,000 for data collection and evaluation. Explain that the remaining $362,000 will come from committed public workforce funds, unrestricted operating revenue, and pending corporate grants. Do not imply pending funds are secured.\n- Sustainability: We will use completion, certification, placement, and retention evidence to pursue renewed public workforce contracts and healthcare-employer contributions. We will also test a cost-sharing model with employer partners, but no partner has committed to it yet.\n- Funder priorities: economic mobility, equitable access to careers, measurable outcomes, and removal of practical barriers.\n\nWrite polished, specific prose suitable for direct submission. Keep the tone confident but evidence-based, distinguish outputs from outcomes, make the logic between funding and results explicit, and do not use bullet points or invent facts."
                }
            ]
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "textInference",
      "taskUUID": "4e2bba2d-8fd1-4a14-bf4e-6686272ae679",
      "model": "openai:gpt@5-mini",
      "settings": {
        "systemPrompt": "You are a senior nonprofit grant writer. Produce persuasive, funder-ready prose grounded strictly in the supplied facts. Never invent statistics, partnerships, commitments, or evaluation results. Preserve requested headings and limits, connect activities to measurable outcomes, and use clear language rather than inflated claims.",
        "maxTokens": 5000,
        "thinkingLevel": "medium"
      },
      "messages": [
        {
          "role": "user",
          "content": "I need help drafting a grant request for Harbor Skills Alliance, a workforce nonprofit in Baltimore. We are seeking support to expand our certified nursing assistant training program. What information do you need?"
        },
        {
          "role": "assistant",
          "content": "Please provide the funding request and grant period; community need; target population; program activities and timeline; recruitment and employer partners; projected outputs and outcomes; evaluation methods; organizational qualifications; total project budget and use of grant funds; and the funder's priorities, required headings, and word limit. I will avoid unsupported claims and mark any essential missing information."
        },
        {
          "role": "user",
          "content": "Draft a complete grant narrative of 1,200–1,400 words for the Northstar Community Foundation under these exact headings: Community Need; Program Design; Outcomes and Evaluation; Organizational Capacity; Use of Funds; Sustainability. End with a concise two-sentence closing case for support.\n\nFacts:\n- Applicant: Harbor Skills Alliance, a Baltimore 501(c)(3) founded in 2014.\n- Request: $250,000 for a 12-month expansion of Pathways to Care, our certified nursing assistant training and placement program.\n- People served: 120 Baltimore City residents, primarily unemployed or underemployed adults ages 20–45. Recruitment will focus on residents of ZIP codes 21215, 21217, and 21223. Participants must be legally eligible to work and able to meet Maryland CNA training requirements.\n- Need: In the last program year, 438 people applied for 72 available seats. Do not add external labor-market statistics; explain the need using only this demand and the barriers listed below.\n- Participant barriers identified during intake: training tuition, transportation, dependent-care costs, limited experience navigating healthcare hiring, and loss of income while attending class.\n- Program model: Four cohorts of 30 participants. Each cohort receives eight weeks of state-approved CNA instruction delivered by Baltimore Allied Health Institute, followed by four weeks of paid clinical placement and structured job-search support.\n- Support: Tuition and exam fees are fully covered. Participants receive a $150 weekly training stipend for 12 weeks, an unlimited local transit pass, dependent-care assistance of up to $800, loaner scrubs, and exam-preparation tutoring.\n- Career services: Individual career plan, résumé and interview coaching, two employer hiring events per cohort, application assistance, and 90 days of post-placement retention coaching.\n- Confirmed employer partners: Harborview Senior Living, Eastgate Rehabilitation Center, and St. Anne's Home Health. Describe them only as confirmed recruitment and interview partners; do not promise hires.\n- Targets: 120 enrolled; at least 102 complete training; at least 96 take the Maryland CNA exam; at least 86 earn certification; at least 78 obtain CNA employment within 90 days of certification; at least 66 of those placed remain employed for 90 days.\n- Evaluation: Attendance records, training completion records, exam registration and results, participant support logs, employer-verified placement, and 30-, 60-, and 90-day retention checks. Program director reviews a dashboard monthly; leadership reviews results quarterly. Participant surveys are administered at enrollment, completion, and 90 days after placement.\n- Capacity: 11 full-time staff. Pathways to Care launched in 2019 and has served 310 participants to date. Current team includes a program director, two career coaches, and a benefits navigator. Baltimore Allied Health Institute supplies licensed instructors and the approved curriculum. The organization has managed state and private grants up to $600,000.\n- Project budget: $612,000 total. Northstar funds will support $96,000 in participant stipends, $48,000 in tuition and exam fees, $36,000 in transportation and dependent-care assistance, $42,000 for one additional career coach, $18,000 for tutoring and training supplies, and $10,000 for data collection and evaluation. Explain that the remaining $362,000 will come from committed public workforce funds, unrestricted operating revenue, and pending corporate grants. Do not imply pending funds are secured.\n- Sustainability: We will use completion, certification, placement, and retention evidence to pursue renewed public workforce contracts and healthcare-employer contributions. We will also test a cost-sharing model with employer partners, but no partner has committed to it yet.\n- Funder priorities: economic mobility, equitable access to careers, measurable outcomes, and removal of practical barriers.\n\nWrite polished, specific prose suitable for direct submission. Keep the tone confident but evidence-based, distinguish outputs from outcomes, make the logic between funding and results explicit, and do not use bullet points or invent facts."
        }
      ]
    }
  ]'

runware run openai:gpt@5-mini \
  settings.systemPrompt="You are a senior nonprofit grant writer. Produce persuasive, funder-ready prose grounded strictly in the supplied facts. Never invent statistics, partnerships, commitments, or evaluation results. Preserve requested headings and limits, connect activities to measurable outcomes, and use clear language rather than inflated claims." \
  settings.maxTokens=5000 \
  settings.thinkingLevel=medium \
  messages.0.role=user \
  messages.0.content="I need help drafting a grant request for Harbor Skills Alliance, a workforce nonprofit in Baltimore. We are seeking support to expand our certified nursing assistant training program. What information do you need?" \
  messages.1.role=assistant \
  messages.1.content="Please provide the funding request and grant period; community need; target population; program activities and timeline; recruitment and employer partners; projected outputs and outcomes; evaluation methods; organizational qualifications; total project budget and use of grant funds; and the funder's priorities, required headings, and word limit. I will avoid unsupported claims and mark any essential missing information." \
  messages.2.role=user \
  messages.2.content="Draft a complete grant narrative of 1,200–1,400 words for the Northstar Community Foundation under these exact headings: Community Need; Program Design; Outcomes and Evaluation; Organizational Capacity; Use of Funds; Sustainability. End with a concise two-sentence closing case for support.

Facts:
- Applicant: Harbor Skills Alliance, a Baltimore 501(c)(3) founded in 2014.
- Request: \$250,000 for a 12-month expansion of Pathways to Care, our certified nursing assistant training and placement program.
- People served: 120 Baltimore City residents, primarily unemployed or underemployed adults ages 20–45. Recruitment will focus on residents of ZIP codes 21215, 21217, and 21223. Participants must be legally eligible to work and able to meet Maryland CNA training requirements.
- Need: In the last program year, 438 people applied for 72 available seats. Do not add external labor-market statistics; explain the need using only this demand and the barriers listed below.
- Participant barriers identified during intake: training tuition, transportation, dependent-care costs, limited experience navigating healthcare hiring, and loss of income while attending class.
- Program model: Four cohorts of 30 participants. Each cohort receives eight weeks of state-approved CNA instruction delivered by Baltimore Allied Health Institute, followed by four weeks of paid clinical placement and structured job-search support.
- Support: Tuition and exam fees are fully covered. Participants receive a \$150 weekly training stipend for 12 weeks, an unlimited local transit pass, dependent-care assistance of up to \$800, loaner scrubs, and exam-preparation tutoring.
- Career services: Individual career plan, résumé and interview coaching, two employer hiring events per cohort, application assistance, and 90 days of post-placement retention coaching.
- Confirmed employer partners: Harborview Senior Living, Eastgate Rehabilitation Center, and St. Anne's Home Health. Describe them only as confirmed recruitment and interview partners; do not promise hires.
- Targets: 120 enrolled; at least 102 complete training; at least 96 take the Maryland CNA exam; at least 86 earn certification; at least 78 obtain CNA employment within 90 days of certification; at least 66 of those placed remain employed for 90 days.
- Evaluation: Attendance records, training completion records, exam registration and results, participant support logs, employer-verified placement, and 30-, 60-, and 90-day retention checks. Program director reviews a dashboard monthly; leadership reviews results quarterly. Participant surveys are administered at enrollment, completion, and 90 days after placement.
- Capacity: 11 full-time staff. Pathways to Care launched in 2019 and has served 310 participants to date. Current team includes a program director, two career coaches, and a benefits navigator. Baltimore Allied Health Institute supplies licensed instructors and the approved curriculum. The organization has managed state and private grants up to \$600,000.
- Project budget: \$612,000 total. Northstar funds will support \$96,000 in participant stipends, \$48,000 in tuition and exam fees, \$36,000 in transportation and dependent-care assistance, \$42,000 for one additional career coach, \$18,000 for tutoring and training supplies, and \$10,000 for data collection and evaluation. Explain that the remaining \$362,000 will come from committed public workforce funds, unrestricted operating revenue, and pending corporate grants. Do not imply pending funds are secured.
- Sustainability: We will use completion, certification, placement, and retention evidence to pursue renewed public workforce contracts and healthcare-employer contributions. We will also test a cost-sharing model with employer partners, but no partner has committed to it yet.
- Funder priorities: economic mobility, equitable access to careers, measurable outcomes, and removal of practical barriers.

Write polished, specific prose suitable for direct submission. Keep the tone confident but evidence-based, distinguish outputs from outcomes, make the logic between funding and results explicit, and do not use bullet points or invent facts."

{
  "taskType": "textInference",
  "taskUUID": "4e2bba2d-8fd1-4a14-bf4e-6686272ae679",
  "model": "openai:gpt@5-mini",
  "settings": {
    "systemPrompt": "You are a senior nonprofit grant writer. Produce persuasive, funder-ready prose grounded strictly in the supplied facts. Never invent statistics, partnerships, commitments, or evaluation results. Preserve requested headings and limits, connect activities to measurable outcomes, and use clear language rather than inflated claims.",
    "maxTokens": 5000,
    "thinkingLevel": "medium"
  },
  "messages": [
    {
      "role": "user",
      "content": "I need help drafting a grant request for Harbor Skills Alliance, a workforce nonprofit in Baltimore. We are seeking support to expand our certified nursing assistant training program. What information do you need?"
    },
    {
      "role": "assistant",
      "content": "Please provide the funding request and grant period; community need; target population; program activities and timeline; recruitment and employer partners; projected outputs and outcomes; evaluation methods; organizational qualifications; total project budget and use of grant funds; and the funder's priorities, required headings, and word limit. I will avoid unsupported claims and mark any essential missing information."
    },
    {
      "role": "user",
      "content": "Draft a complete grant narrative of 1,200–1,400 words for the Northstar Community Foundation under these exact headings: Community Need; Program Design; Outcomes and Evaluation; Organizational Capacity; Use of Funds; Sustainability. End with a concise two-sentence closing case for support.\n\nFacts:\n- Applicant: Harbor Skills Alliance, a Baltimore 501(c)(3) founded in 2014.\n- Request: $250,000 for a 12-month expansion of Pathways to Care, our certified nursing assistant training and placement program.\n- People served: 120 Baltimore City residents, primarily unemployed or underemployed adults ages 20–45. Recruitment will focus on residents of ZIP codes 21215, 21217, and 21223. Participants must be legally eligible to work and able to meet Maryland CNA training requirements.\n- Need: In the last program year, 438 people applied for 72 available seats. Do not add external labor-market statistics; explain the need using only this demand and the barriers listed below.\n- Participant barriers identified during intake: training tuition, transportation, dependent-care costs, limited experience navigating healthcare hiring, and loss of income while attending class.\n- Program model: Four cohorts of 30 participants. Each cohort receives eight weeks of state-approved CNA instruction delivered by Baltimore Allied Health Institute, followed by four weeks of paid clinical placement and structured job-search support.\n- Support: Tuition and exam fees are fully covered. Participants receive a $150 weekly training stipend for 12 weeks, an unlimited local transit pass, dependent-care assistance of up to $800, loaner scrubs, and exam-preparation tutoring.\n- Career services: Individual career plan, résumé and interview coaching, two employer hiring events per cohort, application assistance, and 90 days of post-placement retention coaching.\n- Confirmed employer partners: Harborview Senior Living, Eastgate Rehabilitation Center, and St. Anne's Home Health. Describe them only as confirmed recruitment and interview partners; do not promise hires.\n- Targets: 120 enrolled; at least 102 complete training; at least 96 take the Maryland CNA exam; at least 86 earn certification; at least 78 obtain CNA employment within 90 days of certification; at least 66 of those placed remain employed for 90 days.\n- Evaluation: Attendance records, training completion records, exam registration and results, participant support logs, employer-verified placement, and 30-, 60-, and 90-day retention checks. Program director reviews a dashboard monthly; leadership reviews results quarterly. Participant surveys are administered at enrollment, completion, and 90 days after placement.\n- Capacity: 11 full-time staff. Pathways to Care launched in 2019 and has served 310 participants to date. Current team includes a program director, two career coaches, and a benefits navigator. Baltimore Allied Health Institute supplies licensed instructors and the approved curriculum. The organization has managed state and private grants up to $600,000.\n- Project budget: $612,000 total. Northstar funds will support $96,000 in participant stipends, $48,000 in tuition and exam fees, $36,000 in transportation and dependent-care assistance, $42,000 for one additional career coach, $18,000 for tutoring and training supplies, and $10,000 for data collection and evaluation. Explain that the remaining $362,000 will come from committed public workforce funds, unrestricted operating revenue, and pending corporate grants. Do not imply pending funds are secured.\n- Sustainability: We will use completion, certification, placement, and retention evidence to pursue renewed public workforce contracts and healthcare-employer contributions. We will also test a cost-sharing model with employer partners, but no partner has committed to it yet.\n- Funder priorities: economic mobility, equitable access to careers, measurable outcomes, and removal of practical barriers.\n\nWrite polished, specific prose suitable for direct submission. Keep the tone confident but evidence-based, distinguish outputs from outcomes, make the logic between funding and results explicit, and do not use bullet points or invent facts."
    }
  ]
}

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'openai:gpt@5-mini',
  settings: {
    maxTokens: 3500,
    thinkingLevel: 'medium'
  },
  inputs: {
    images: [
      'https://assets.runware.ai/assets/inputs/9a41a1db-675f-44a6-ab32-886def7e2032.jpg'
    ]
  },
  messages: [
    {
      role: 'user',
      content: 'I manage a 28,000-square-foot retail property and need a preliminary roof assessment based on the attached drone photograph. Prepare a practical package I can send to our property director.\n\nInclude:\n1. An executive summary of the roof\'s apparent condition.\n2. A Markdown table of visible findings with location, visual evidence, likely concern, urgency level, and recommended next step.\n3. A prioritized 30-day action plan divided into immediate, within 7 days, and within 30 days.\n4. A concise scope of work that I can use to request quotes from licensed commercial roofing contractors.\n5. A short email to the property director summarizing the issue and requesting authorization for an on-site inspection.\n\nSeparate direct visual observations from interpretations. Do not invent measurements, leak history, membrane age, code violations, costs, or conditions that cannot be established from the image. Use calibrated language such as \'appears,\' \'may,\' and \'cannot be confirmed from the photograph.\' Flag any conditions that could present an immediate safety or water-intrusion risk, and state clearly that this desktop review does not replace an on-site inspection by a qualified roofing professional.'
    }
  ]
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "openai:gpt@5-mini",
            "settings": {
                "maxTokens": 3500,
                "thinkingLevel": "medium"
            },
            "inputs": {
                "images": [
                    "https://assets.runware.ai/assets/inputs/9a41a1db-675f-44a6-ab32-886def7e2032.jpg"
                ]
            },
            "messages": [
                {
                    "role": "user",
                    "content": "I manage a 28,000-square-foot retail property and need a preliminary roof assessment based on the attached drone photograph. Prepare a practical package I can send to our property director.\n\nInclude:\n1. An executive summary of the roof's apparent condition.\n2. A Markdown table of visible findings with location, visual evidence, likely concern, urgency level, and recommended next step.\n3. A prioritized 30-day action plan divided into immediate, within 7 days, and within 30 days.\n4. A concise scope of work that I can use to request quotes from licensed commercial roofing contractors.\n5. A short email to the property director summarizing the issue and requesting authorization for an on-site inspection.\n\nSeparate direct visual observations from interpretations. Do not invent measurements, leak history, membrane age, code violations, costs, or conditions that cannot be established from the image. Use calibrated language such as 'appears,' 'may,' and 'cannot be confirmed from the photograph.' Flag any conditions that could present an immediate safety or water-intrusion risk, and state clearly that this desktop review does not replace an on-site inspection by a qualified roofing professional."
                }
            ]
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "textInference",
      "taskUUID": "9da4cddd-ecfc-4b9d-9062-0796a6de70f6",
      "model": "openai:gpt@5-mini",
      "settings": {
        "maxTokens": 3500,
        "thinkingLevel": "medium"
      },
      "inputs": {
        "images": [
          "https://assets.runware.ai/assets/inputs/9a41a1db-675f-44a6-ab32-886def7e2032.jpg"
        ]
      },
      "messages": [
        {
          "role": "user",
          "content": "I manage a 28,000-square-foot retail property and need a preliminary roof assessment based on the attached drone photograph. Prepare a practical package I can send to our property director.\n\nInclude:\n1. An executive summary of the roof's apparent condition.\n2. A Markdown table of visible findings with location, visual evidence, likely concern, urgency level, and recommended next step.\n3. A prioritized 30-day action plan divided into immediate, within 7 days, and within 30 days.\n4. A concise scope of work that I can use to request quotes from licensed commercial roofing contractors.\n5. A short email to the property director summarizing the issue and requesting authorization for an on-site inspection.\n\nSeparate direct visual observations from interpretations. Do not invent measurements, leak history, membrane age, code violations, costs, or conditions that cannot be established from the image. Use calibrated language such as 'appears,' 'may,' and 'cannot be confirmed from the photograph.' Flag any conditions that could present an immediate safety or water-intrusion risk, and state clearly that this desktop review does not replace an on-site inspection by a qualified roofing professional."
        }
      ]
    }
  ]'

runware run openai:gpt@5-mini \
  settings.maxTokens=3500 \
  settings.thinkingLevel=medium \
  inputs.images.0=https://assets.runware.ai/assets/inputs/9a41a1db-675f-44a6-ab32-886def7e2032.jpg \
  messages.0.role=user \
  messages.0.content="I manage a 28,000-square-foot retail property and need a preliminary roof assessment based on the attached drone photograph. Prepare a practical package I can send to our property director.

Include:
1. An executive summary of the roof's apparent condition.
2. A Markdown table of visible findings with location, visual evidence, likely concern, urgency level, and recommended next step.
3. A prioritized 30-day action plan divided into immediate, within 7 days, and within 30 days.
4. A concise scope of work that I can use to request quotes from licensed commercial roofing contractors.
5. A short email to the property director summarizing the issue and requesting authorization for an on-site inspection.

Separate direct visual observations from interpretations. Do not invent measurements, leak history, membrane age, code violations, costs, or conditions that cannot be established from the image. Use calibrated language such as 'appears,' 'may,' and 'cannot be confirmed from the photograph.' Flag any conditions that could present an immediate safety or water-intrusion risk, and state clearly that this desktop review does not replace an on-site inspection by a qualified roofing professional."

{
  "taskType": "textInference",
  "taskUUID": "9da4cddd-ecfc-4b9d-9062-0796a6de70f6",
  "model": "openai:gpt@5-mini",
  "settings": {
    "maxTokens": 3500,
    "thinkingLevel": "medium"
  },
  "inputs": {
    "images": [
      "https://assets.runware.ai/assets/inputs/9a41a1db-675f-44a6-ab32-886def7e2032.jpg"
    ]
  },
  "messages": [
    {
      "role": "user",
      "content": "I manage a 28,000-square-foot retail property and need a preliminary roof assessment based on the attached drone photograph. Prepare a practical package I can send to our property director.\n\nInclude:\n1. An executive summary of the roof's apparent condition.\n2. A Markdown table of visible findings with location, visual evidence, likely concern, urgency level, and recommended next step.\n3. A prioritized 30-day action plan divided into immediate, within 7 days, and within 30 days.\n4. A concise scope of work that I can use to request quotes from licensed commercial roofing contractors.\n5. A short email to the property director summarizing the issue and requesting authorization for an on-site inspection.\n\nSeparate direct visual observations from interpretations. Do not invent measurements, leak history, membrane age, code violations, costs, or conditions that cannot be established from the image. Use calibrated language such as 'appears,' 'may,' and 'cannot be confirmed from the photograph.' Flag any conditions that could present an immediate safety or water-intrusion risk, and state clearly that this desktop review does not replace an on-site inspection by a qualified roofing professional."
    }
  ]
}

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'openai:gpt@5-mini',
  settings: {
    systemPrompt: 'You are a senior incident communications editor for a B2B SaaS company. Convert technical notes into clear, accountable customer communications. Preserve every supplied fact, never invent causes or commitments, distinguish confirmed findings from ongoing work, use plain language, and avoid minimizing customer impact. Return polished Markdown ready for publication.',
    maxTokens: 5000,
    thinkingLevel: 'medium'
  },
  messages: [
    {
      role: 'user',
      content: 'Create the complete incident communications package for Northstar Metrics, an analytics platform used by retail operations teams.\n\nDeliver exactly these sections:\n1. Final Status-Page Update — 120–180 words\n2. Customer Email — subject line plus 250–350 words\n3. Post-Incident Report — 700–1,000 words with the headings Summary, Customer Impact, Timeline, Root Cause, Resolution and Recovery, Corrective Actions\n\nAudience: operations leaders and technical administrators. Tone: direct, calm, transparent, and accountable. Explain technical details in accessible language. Use UTC for every timestamp. Do not claim data loss, security exposure, or SLA eligibility unless the notes establish it. Do not add dates, metrics, causes, or deadlines. In Corrective Actions, clearly separate completed work from planned work.\n\nInternal engineering notes:\n- Incident date: 18 July 2026\n- Customer-facing incident window: 14:07–15:41 UTC\n- At 14:07, dashboard queries began timing out for customers whose workspaces were located in the EU region.\n- API ingestion remained available throughout. Incoming events were accepted and queued, so dashboards displayed delayed data after service returned.\n- US-region workspaces were not affected.\n- Peak impact: 38% of EU dashboard query requests failed; remaining requests were slower than normal.\n- First automated alert fired at 14:10 for elevated query latency.\n- On-call engineer acknowledged at 14:13.\n- Incident declared at 14:18.\n- Status page posted at 14:24.\n- Engineers identified saturation in the EU query cluster at 14:31.\n- Immediate trigger: a routine configuration deployment at 14:02 changed the cache connection pool limit from 800 to 80 because a zero was omitted in the region-specific value.\n- The deployment passed automated validation because validation checked that the value was a positive integer but did not compare it with the prior value or expected operating range.\n- Connection contention caused queries to queue, retry, and amplify load on the cluster.\n- Rollback began at 14:36 and completed at 14:44.\n- Error rates fell after rollback, but a retry backlog kept latency elevated.\n- Engineers temporarily added query capacity at 14:52 and drained the backlog.\n- Dashboard query success rate returned to normal at 15:22.\n- Monitoring continued until 15:41, when the incident was resolved.\n- All queued ingestion data was processed by 16:05.\n- Confirmed: no event data was lost.\n- The investigation found no indication of unauthorized access or exposure of customer data.\n- Completed corrective actions: restored the intended pool limit; added an operating-range check for cache connection settings; added an alert for sudden regional connection-pool changes; required a second engineer to approve production configuration changes.\n- Planned corrective actions: create automated comparison of regional configuration values; add a pre-deployment load simulation for query-cluster settings; run an incident-response exercise focused on regional partial outages.\n- No completion dates have been approved for planned actions.\n- Avoid the phrase “brief disruption.”\n- Refer to the product as Northstar Metrics on first mention and “the platform” afterward.'
    }
  ]
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "openai:gpt@5-mini",
            "settings": {
                "systemPrompt": "You are a senior incident communications editor for a B2B SaaS company. Convert technical notes into clear, accountable customer communications. Preserve every supplied fact, never invent causes or commitments, distinguish confirmed findings from ongoing work, use plain language, and avoid minimizing customer impact. Return polished Markdown ready for publication.",
                "maxTokens": 5000,
                "thinkingLevel": "medium"
            },
            "messages": [
                {
                    "role": "user",
                    "content": "Create the complete incident communications package for Northstar Metrics, an analytics platform used by retail operations teams.\n\nDeliver exactly these sections:\n1. Final Status-Page Update — 120–180 words\n2. Customer Email — subject line plus 250–350 words\n3. Post-Incident Report — 700–1,000 words with the headings Summary, Customer Impact, Timeline, Root Cause, Resolution and Recovery, Corrective Actions\n\nAudience: operations leaders and technical administrators. Tone: direct, calm, transparent, and accountable. Explain technical details in accessible language. Use UTC for every timestamp. Do not claim data loss, security exposure, or SLA eligibility unless the notes establish it. Do not add dates, metrics, causes, or deadlines. In Corrective Actions, clearly separate completed work from planned work.\n\nInternal engineering notes:\n- Incident date: 18 July 2026\n- Customer-facing incident window: 14:07–15:41 UTC\n- At 14:07, dashboard queries began timing out for customers whose workspaces were located in the EU region.\n- API ingestion remained available throughout. Incoming events were accepted and queued, so dashboards displayed delayed data after service returned.\n- US-region workspaces were not affected.\n- Peak impact: 38% of EU dashboard query requests failed; remaining requests were slower than normal.\n- First automated alert fired at 14:10 for elevated query latency.\n- On-call engineer acknowledged at 14:13.\n- Incident declared at 14:18.\n- Status page posted at 14:24.\n- Engineers identified saturation in the EU query cluster at 14:31.\n- Immediate trigger: a routine configuration deployment at 14:02 changed the cache connection pool limit from 800 to 80 because a zero was omitted in the region-specific value.\n- The deployment passed automated validation because validation checked that the value was a positive integer but did not compare it with the prior value or expected operating range.\n- Connection contention caused queries to queue, retry, and amplify load on the cluster.\n- Rollback began at 14:36 and completed at 14:44.\n- Error rates fell after rollback, but a retry backlog kept latency elevated.\n- Engineers temporarily added query capacity at 14:52 and drained the backlog.\n- Dashboard query success rate returned to normal at 15:22.\n- Monitoring continued until 15:41, when the incident was resolved.\n- All queued ingestion data was processed by 16:05.\n- Confirmed: no event data was lost.\n- The investigation found no indication of unauthorized access or exposure of customer data.\n- Completed corrective actions: restored the intended pool limit; added an operating-range check for cache connection settings; added an alert for sudden regional connection-pool changes; required a second engineer to approve production configuration changes.\n- Planned corrective actions: create automated comparison of regional configuration values; add a pre-deployment load simulation for query-cluster settings; run an incident-response exercise focused on regional partial outages.\n- No completion dates have been approved for planned actions.\n- Avoid the phrase “brief disruption.”\n- Refer to the product as Northstar Metrics on first mention and “the platform” afterward."
                }
            ]
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "textInference",
      "taskUUID": "51f316fb-88f6-4cd3-ab17-a524dad2a219",
      "model": "openai:gpt@5-mini",
      "settings": {
        "systemPrompt": "You are a senior incident communications editor for a B2B SaaS company. Convert technical notes into clear, accountable customer communications. Preserve every supplied fact, never invent causes or commitments, distinguish confirmed findings from ongoing work, use plain language, and avoid minimizing customer impact. Return polished Markdown ready for publication.",
        "maxTokens": 5000,
        "thinkingLevel": "medium"
      },
      "messages": [
        {
          "role": "user",
          "content": "Create the complete incident communications package for Northstar Metrics, an analytics platform used by retail operations teams.\n\nDeliver exactly these sections:\n1. Final Status-Page Update — 120–180 words\n2. Customer Email — subject line plus 250–350 words\n3. Post-Incident Report — 700–1,000 words with the headings Summary, Customer Impact, Timeline, Root Cause, Resolution and Recovery, Corrective Actions\n\nAudience: operations leaders and technical administrators. Tone: direct, calm, transparent, and accountable. Explain technical details in accessible language. Use UTC for every timestamp. Do not claim data loss, security exposure, or SLA eligibility unless the notes establish it. Do not add dates, metrics, causes, or deadlines. In Corrective Actions, clearly separate completed work from planned work.\n\nInternal engineering notes:\n- Incident date: 18 July 2026\n- Customer-facing incident window: 14:07–15:41 UTC\n- At 14:07, dashboard queries began timing out for customers whose workspaces were located in the EU region.\n- API ingestion remained available throughout. Incoming events were accepted and queued, so dashboards displayed delayed data after service returned.\n- US-region workspaces were not affected.\n- Peak impact: 38% of EU dashboard query requests failed; remaining requests were slower than normal.\n- First automated alert fired at 14:10 for elevated query latency.\n- On-call engineer acknowledged at 14:13.\n- Incident declared at 14:18.\n- Status page posted at 14:24.\n- Engineers identified saturation in the EU query cluster at 14:31.\n- Immediate trigger: a routine configuration deployment at 14:02 changed the cache connection pool limit from 800 to 80 because a zero was omitted in the region-specific value.\n- The deployment passed automated validation because validation checked that the value was a positive integer but did not compare it with the prior value or expected operating range.\n- Connection contention caused queries to queue, retry, and amplify load on the cluster.\n- Rollback began at 14:36 and completed at 14:44.\n- Error rates fell after rollback, but a retry backlog kept latency elevated.\n- Engineers temporarily added query capacity at 14:52 and drained the backlog.\n- Dashboard query success rate returned to normal at 15:22.\n- Monitoring continued until 15:41, when the incident was resolved.\n- All queued ingestion data was processed by 16:05.\n- Confirmed: no event data was lost.\n- The investigation found no indication of unauthorized access or exposure of customer data.\n- Completed corrective actions: restored the intended pool limit; added an operating-range check for cache connection settings; added an alert for sudden regional connection-pool changes; required a second engineer to approve production configuration changes.\n- Planned corrective actions: create automated comparison of regional configuration values; add a pre-deployment load simulation for query-cluster settings; run an incident-response exercise focused on regional partial outages.\n- No completion dates have been approved for planned actions.\n- Avoid the phrase “brief disruption.”\n- Refer to the product as Northstar Metrics on first mention and “the platform” afterward."
        }
      ]
    }
  ]'

runware run openai:gpt@5-mini \
  settings.systemPrompt="You are a senior incident communications editor for a B2B SaaS company. Convert technical notes into clear, accountable customer communications. Preserve every supplied fact, never invent causes or commitments, distinguish confirmed findings from ongoing work, use plain language, and avoid minimizing customer impact. Return polished Markdown ready for publication." \
  settings.maxTokens=5000 \
  settings.thinkingLevel=medium \
  messages.0.role=user \
  messages.0.content="Create the complete incident communications package for Northstar Metrics, an analytics platform used by retail operations teams.

Deliver exactly these sections:
1. Final Status-Page Update — 120–180 words
2. Customer Email — subject line plus 250–350 words
3. Post-Incident Report — 700–1,000 words with the headings Summary, Customer Impact, Timeline, Root Cause, Resolution and Recovery, Corrective Actions

Audience: operations leaders and technical administrators. Tone: direct, calm, transparent, and accountable. Explain technical details in accessible language. Use UTC for every timestamp. Do not claim data loss, security exposure, or SLA eligibility unless the notes establish it. Do not add dates, metrics, causes, or deadlines. In Corrective Actions, clearly separate completed work from planned work.

Internal engineering notes:
- Incident date: 18 July 2026
- Customer-facing incident window: 14:07–15:41 UTC
- At 14:07, dashboard queries began timing out for customers whose workspaces were located in the EU region.
- API ingestion remained available throughout. Incoming events were accepted and queued, so dashboards displayed delayed data after service returned.
- US-region workspaces were not affected.
- Peak impact: 38% of EU dashboard query requests failed; remaining requests were slower than normal.
- First automated alert fired at 14:10 for elevated query latency.
- On-call engineer acknowledged at 14:13.
- Incident declared at 14:18.
- Status page posted at 14:24.
- Engineers identified saturation in the EU query cluster at 14:31.
- Immediate trigger: a routine configuration deployment at 14:02 changed the cache connection pool limit from 800 to 80 because a zero was omitted in the region-specific value.
- The deployment passed automated validation because validation checked that the value was a positive integer but did not compare it with the prior value or expected operating range.
- Connection contention caused queries to queue, retry, and amplify load on the cluster.
- Rollback began at 14:36 and completed at 14:44.
- Error rates fell after rollback, but a retry backlog kept latency elevated.
- Engineers temporarily added query capacity at 14:52 and drained the backlog.
- Dashboard query success rate returned to normal at 15:22.
- Monitoring continued until 15:41, when the incident was resolved.
- All queued ingestion data was processed by 16:05.
- Confirmed: no event data was lost.
- The investigation found no indication of unauthorized access or exposure of customer data.
- Completed corrective actions: restored the intended pool limit; added an operating-range check for cache connection settings; added an alert for sudden regional connection-pool changes; required a second engineer to approve production configuration changes.
- Planned corrective actions: create automated comparison of regional configuration values; add a pre-deployment load simulation for query-cluster settings; run an incident-response exercise focused on regional partial outages.
- No completion dates have been approved for planned actions.
- Avoid the phrase “brief disruption.”
- Refer to the product as Northstar Metrics on first mention and “the platform” afterward."

{
  "taskType": "textInference",
  "taskUUID": "51f316fb-88f6-4cd3-ab17-a524dad2a219",
  "model": "openai:gpt@5-mini",
  "settings": {
    "systemPrompt": "You are a senior incident communications editor for a B2B SaaS company. Convert technical notes into clear, accountable customer communications. Preserve every supplied fact, never invent causes or commitments, distinguish confirmed findings from ongoing work, use plain language, and avoid minimizing customer impact. Return polished Markdown ready for publication.",
    "maxTokens": 5000,
    "thinkingLevel": "medium"
  },
  "messages": [
    {
      "role": "user",
      "content": "Create the complete incident communications package for Northstar Metrics, an analytics platform used by retail operations teams.\n\nDeliver exactly these sections:\n1. Final Status-Page Update — 120–180 words\n2. Customer Email — subject line plus 250–350 words\n3. Post-Incident Report — 700–1,000 words with the headings Summary, Customer Impact, Timeline, Root Cause, Resolution and Recovery, Corrective Actions\n\nAudience: operations leaders and technical administrators. Tone: direct, calm, transparent, and accountable. Explain technical details in accessible language. Use UTC for every timestamp. Do not claim data loss, security exposure, or SLA eligibility unless the notes establish it. Do not add dates, metrics, causes, or deadlines. In Corrective Actions, clearly separate completed work from planned work.\n\nInternal engineering notes:\n- Incident date: 18 July 2026\n- Customer-facing incident window: 14:07–15:41 UTC\n- At 14:07, dashboard queries began timing out for customers whose workspaces were located in the EU region.\n- API ingestion remained available throughout. Incoming events were accepted and queued, so dashboards displayed delayed data after service returned.\n- US-region workspaces were not affected.\n- Peak impact: 38% of EU dashboard query requests failed; remaining requests were slower than normal.\n- First automated alert fired at 14:10 for elevated query latency.\n- On-call engineer acknowledged at 14:13.\n- Incident declared at 14:18.\n- Status page posted at 14:24.\n- Engineers identified saturation in the EU query cluster at 14:31.\n- Immediate trigger: a routine configuration deployment at 14:02 changed the cache connection pool limit from 800 to 80 because a zero was omitted in the region-specific value.\n- The deployment passed automated validation because validation checked that the value was a positive integer but did not compare it with the prior value or expected operating range.\n- Connection contention caused queries to queue, retry, and amplify load on the cluster.\n- Rollback began at 14:36 and completed at 14:44.\n- Error rates fell after rollback, but a retry backlog kept latency elevated.\n- Engineers temporarily added query capacity at 14:52 and drained the backlog.\n- Dashboard query success rate returned to normal at 15:22.\n- Monitoring continued until 15:41, when the incident was resolved.\n- All queued ingestion data was processed by 16:05.\n- Confirmed: no event data was lost.\n- The investigation found no indication of unauthorized access or exposure of customer data.\n- Completed corrective actions: restored the intended pool limit; added an operating-range check for cache connection settings; added an alert for sudden regional connection-pool changes; required a second engineer to approve production configuration changes.\n- Planned corrective actions: create automated comparison of regional configuration values; add a pre-deployment load simulation for query-cluster settings; run an incident-response exercise focused on regional partial outages.\n- No completion dates have been approved for planned actions.\n- Avoid the phrase “brief disruption.”\n- Refer to the product as Northstar Metrics on first mention and “the platform” afterward."
    }
  ]
}