Work prompts

Source: app.asar of the Codex/ChatGPT desktop app 26.924.22138 (build 11645), SHA-256 d0ba973179d2f717affd39e012b64a095464a54a51c6bccb7bc6b3d2a1cfba80.

Messages ChatGPT Work sends or prefills during onboarding: starter tasks, the daily-briefing next step, the writing-style skill setup, write-like-me requests and the browser-extension Side Chat samples.

ChatGPT's own system prompt is not in the app; the servers add it. The phrase "You are ChatGPT" occurs in none of the app's scripts.

Each entry says whether its text is exact (one literal in the bundle) or assembled (literal pieces joined as the app joins them). Entries with a message id or translator note are formatjs messages: the text shown is the English source (defaultMessage), and the app sends the model whatever the user's language translates it to.

Onboarding starters

Starter: personal website

Source: webview/assets/fallback-cards-813d1df097fa.js, offset 8207, SHA-256 99a1ec242ca96dc22fd75f70ba95c3f4cdbddc07efd98e83feab5d998c263b47.

Exact text from the bundle. Message id chatgpt.tpp.onboarding.starter.personal_website.prompt.

Translator note: Prompt submitted when the user selects the personal website starter task in the ChatGPT Work onboarding conversation. It asks ChatGPT to choose a useful website subject from context already available in past conversations, then design, build, and host the site immediately. It must not invent personal details or claim access to unavailable context.

Based on everything you know about me and what we’ve discussed in past conversations, choose the most useful subject for a personal website. If there isn’t enough context to choose honestly, ask me one focused question about what the site should be for. Then design, build, and host it for me. Make confident choices about the content, structure, and visual style you think I’ll like best, and start creating it now.

Starter: manage inbox

Source: webview/assets/fallback-cards-813d1df097fa.js, offset 9936, SHA-256 c6996d589e9368ca25ca3b6a5807521ce92f75c663ed419ea0b502632818fa78.

Exact text from the bundle. Message id chatgpt.tpp.onboarding.starter.manage_inbox.prompt.

Translator note: Prompt submitted when the user selects the inbox management starter task in the ChatGPT Work onboarding conversation. It asks ChatGPT to use or help connect the user's preferred email provider, study three months of sent email to learn how the user writes to different people, create a reusable email-writing skill, and create a daily automation that drafts replies to unread email. It explicitly prohibits sending email without confirmation.

Use my connected email app. If I haven’t connected one, ask which email provider I use and help me connect it before continuing. Then review the emails I’ve sent over the last three months and learn how my writing changes for different people and situations. Create a reusable email-writing skill that captures those styles, patterns, and preferences. Set up a daily automation that reviews my unread emails and drafts a response to each one in the appropriate style. Never send an email without my explicit confirmation.

Starter: personalized presentation

Source: webview/assets/fallback-cards-813d1df097fa.js, offset 13415, SHA-256 3aee7ed7c77e6936ab095df4862457c3088b46716121eb3436edbe6e0f9a2f80.

Exact text from the bundle. Message id chatgpt.tpp.onboarding.starter.personalized_presentation.prompt.

Translator note: Prompt submitted when the user selects the presentation starter task in the ChatGPT Work onboarding conversation. It asks ChatGPT to choose a useful topic grounded in the user's past conversations, create a complete polished slide presentation immediately, and provide the finished file with a short explanation. It must not invent a topic or claim context that is unavailable.

Review our past conversations and choose the topic that would be most useful to turn into a polished slide presentation for me right now. If there isn’t enough context to choose honestly, ask me one focused question about the subject. Then pull together the relevant context, decide on a strong story and structure, write the content, and create the complete presentation with useful visuals. Make confident choices and start creating it now. When it’s ready, give me the finished file and a brief note explaining what you chose and why.

Starter: repeatable-work skill

Source: webview/assets/fallback-cards-813d1df097fa.js, offset 14988, SHA-256 70c95b76c0daf513a45e3e80b3415b9df59146bb6791b797cde5784537f1cf14.

Exact text from the bundle.

Look across my past conversations, including older chats, for a workflow I repeat that spans apps, websites, or files. Use only context you can actually access; never invent conversations or examples. If there is not enough evidence to identify a recurring workflow, first call `request_user_input` to ask one focused question about a workflow I repeat and a real example, then wait for my answer before proposing one. Once a workflow is supported by the available history or my answer, show a section titled ‘Workflow I found’ with a plain-language description of the workflow, real supporting examples (2–3 when available), and the background and decisions the skill would preserve. Do not create the skill, call Skill Creator, or continue to any next step in that response. After the evidence is visible, stop and call `request_user_input` to ask: ‘Is this workflow accurate, and should I turn it into a skill?’ Offer two choices: ‘Yes, create it’ and ‘No, revise it’. Do not ask for confirmation in ordinary chat text. Wait for the tool response before doing anything else.

Only after I explicitly confirm, create it as a reusable skill. Give it a descriptive lowercase kebab-case name in the `your-new-skill` format. Do not stop after saving it. In the same response, show a section titled ‘Try it next’ with a numbered list of exactly 3 concrete requests I could use the skill for in future work. Do not invoke or apply the skill yet. End by asking whether I would like to try one of those three examples, and which one.

Starter selection context

Source: webview/assets/home-ambient-suggestions-content-2014aee9e3b2.js, offset 18725, SHA-256 2fca00df9c58d896f7410a511bbdbd087c417d368a57ea14d46dd7835bf04424.

Assembled from literal pieces in the bundle, joined as the app joins them; <…> marks a value filled in at run time.

For the first turn only, treat the user's starter-prompt selection as the following user request. In that request, first-person pronouns refer to the user:

<…>

This instruction applies only to the initial request. Later user messages may revise or cancel it.

Next step: daily briefing

Source: webview/assets/home-9859b126064a.js, offset 83695, SHA-256 1da19cc520de9aa4f0533e6f5808078f8893f9d356d0881069abee7b6e715c96.

Exact text from the bundle. Message id chatgpt.tpp.onboarding.personalize.next_steps.daily_briefing.prompt.user_request. The same text ships at 2 places in the bundle; the first is shown.

Translator note: User-visible prompt submitted when an end user selects Send me a daily briefing in the ChatGPT Work personalization onboarding conversation. It asks ChatGPT to prepare a recurring daily briefing using connected tools such as email, calendar, and Slack, then show the proposed scheduled automation for approval before activating it. Slack is a product name and must remain unchanged. No markup or placeholders.

Create a daily automation that prepares a concise briefing focused on my connected apps like email, calendar, and Slack. Show me the automation before enabling it.

Writing style

Writing-style skill: connected apps to check

Source: webview/assets/home-ambient-suggestions-content-2014aee9e3b2.js, offset 5449, SHA-256 5081b54dffd7ba85ef9e8d49c1bc45417e5af50973d76932fa7eb5625896d922.

Exact text from the bundle. Message id home.ambientSuggestions.learnWritingStyle.prompt.connectionUnknown.v15.

Translator note: Hidden initial request prepared by the Learn my writing style card while communication app connection state is unavailable. It tells Codex to show a short preamble, use request_user_input and request_plugin_install, report non-sensitive style discoveries instead of implementation mechanics, validate without adding dependencies, and explain that the resulting skills activate automatically. Placeholders {emailApp} and {messageApp} are the account's provider-specific email and messaging app names

Use $skill-creator to create a personal writing-style skill for each selected writing context. Begin with one short user-facing preamble before any tool call so I immediately know that you are checking my communication app setup and will help me connect the apps I choose. Then check whether any communication apps are connected. If none are accessible, call `request_user_input` with exactly one required question: header `Apps`, id `writing_style_sources`, and question `Which apps should I learn your writing style from?`. Offer exactly these three options in order: `{emailApp} + {messageApp} (Recommended)` with description `Learn separate styles for email and {messageApp} messages.`, `{emailApp} only` with description `Learn my email writing style.`, and `{messageApp} only` with description `Learn how I write messages in {messageApp}.`. Do not include an Other option because the input UI adds it automatically. Do not set `autoResolutionMs` because this choice is required. If no selection is returned, do not guess or continue. For each selected app that is not accessible, call `request_plugin_install` one app at a time using its matching `plugin_id` from the recommended plugin list and a concise reason. Do not ask me to connect an app manually before attempting this install flow. Do not search the filesystem, app caches, or credential files for connector state. After each installation and connection completes, recheck app access and continue automatically. If a matching plugin is unavailable, the request is declined, or connection fails, briefly explain the blocker and stop without creating that writing context's skill. Keep progress updates sparse and substantive. Before sampling, limit them to the next user-relevant milestone. Once sampling begins, use them to share one or two non-sensitive style traits identified so far and how the sample is becoming more representative. Do not narrate filesystem operations, helper scripts, dependency checks, or validation mechanics. Do not ask where to install the skills; use the default personal skills directory. Analyze a representative sample of communications authored by me in each connected app. Exclude received text, quoted replies, forwarded content, signatures, automated notifications, and boilerplate. Consider variation across recipients, situations, and time. Capture stable stylistic traits—tone, sentence structure, vocabulary, formatting, level of detail, and recurring patterns—without treating topics, names, or confidential facts as style. If there is not enough evidence, explain what is missing rather than inventing conclusions. Create separate skills for email and {messageApp} messages. Combine equivalent email connectors into one email skill, but do not combine email with {messageApp}. For each writing context, create a concise SKILL.md and a detailed `references/writing-style.md`. Configure the skill to trigger automatically for requests to draft, rewrite, or reply in that context, while avoiding overlap with the other writing-style skills. Do not reproduce or lightly rewrite source communications in generated files or responses. Remove names, sensitive details, and recognizable scenarios. Validate without adding dependencies: confirm the required files exist and that each SKILL.md has valid frontmatter with a hyphen-case name and a non-empty description. Run a bundled validator only if it works unmodified with already-available dependencies. Never install, vendor, or mock validation dependencies, and never claim validator success unless it actually ran. Fix structural or validation failures before reporting completion. In the final response, mention only skills that were successfully created and installed. Include one brief sentence explaining that a skill is a reusable set of instructions that can be automatically applied to matching tasks. Briefly explain that each writing-style skill activates automatically for ordinary requests to draft, rewrite, or reply in its context. Do not suggest that I must name the skill, invoke it explicitly, or say "write in my voice". For each writing context, preview `Overall voice`, `Common patterns`, and `Do and avoid`, followed by 2–3 entirely new example messages formatted as individual Markdown block quotes. End each section with a link to its full writing-style reference.

Writing-style skill: no app connected

Source: webview/assets/home-ambient-suggestions-content-2014aee9e3b2.js, offset 10567, SHA-256 ae1ad24dbfdbc9d903f709eeb4da72e032dc1c2f60d052d741b2511c7a5d2728.

Exact text from the bundle. Message id home.ambientSuggestions.learnWritingStyle.prompt.noConnections.v15.

Translator note: Hidden initial request prepared by the Learn my writing style card when no communication app is connected. It tells Codex to show a short preamble, use request_user_input and request_plugin_install, report non-sensitive style discoveries instead of implementation mechanics, validate without adding dependencies, and explain that the resulting skills activate automatically. Placeholders {emailApp} and {messageApp} are the account's provider-specific email and messaging app names

Use $skill-creator to create a personal writing-style skill for each selected writing context. No communication app is connected yet. Begin with one short user-facing preamble before any tool call so I immediately know that you will ask which communication apps to learn from and help me connect them. Then call `request_user_input` with exactly one required question: header `Apps`, id `writing_style_sources`, and question `Which apps should I learn your writing style from?`. Offer exactly these three options in order: `{emailApp} + {messageApp} (Recommended)` with description `Learn separate styles for email and {messageApp} messages.`, `{emailApp} only` with description `Learn my email writing style.`, and `{messageApp} only` with description `Learn how I write messages in {messageApp}.`. Do not include an Other option because the input UI adds it automatically. Do not set `autoResolutionMs` because this choice is required. If no selection is returned, do not guess or continue. For each selected app that is not accessible, call `request_plugin_install` one app at a time using its matching `plugin_id` from the recommended plugin list and a concise reason. Do not ask me to connect an app manually before attempting this install flow. Do not search the filesystem, app caches, or credential files for connector state. After each installation and connection completes, recheck app access and continue automatically. If a matching plugin is unavailable, the request is declined, or connection fails, briefly explain the blocker and stop without creating that writing context's skill. Keep progress updates sparse and substantive. Before sampling, limit them to the next user-relevant milestone. Once sampling begins, use them to share one or two non-sensitive style traits identified so far and how the sample is becoming more representative. Do not narrate filesystem operations, helper scripts, dependency checks, or validation mechanics. Do not ask where to install the skills; use the default personal skills directory. Analyze a representative sample of communications authored by me in each connected app. Exclude received text, quoted replies, forwarded content, signatures, automated notifications, and boilerplate. Consider variation across recipients, situations, and time. Capture stable stylistic traits—tone, sentence structure, vocabulary, formatting, level of detail, and recurring patterns—without treating topics, names, or confidential facts as style. If there is not enough evidence, explain what is missing rather than inventing conclusions. Create separate skills for email and {messageApp} messages. Combine equivalent email connectors into one email skill, but do not combine email with {messageApp}. For each writing context, create a concise SKILL.md and a detailed `references/writing-style.md`. Configure the skill to trigger automatically for requests to draft, rewrite, or reply in that context, while avoiding overlap with the other writing-style skills. Do not reproduce or lightly rewrite source communications in generated files or responses. Remove names, sensitive details, and recognizable scenarios. Validate without adding dependencies: confirm the required files exist and that each SKILL.md has valid frontmatter with a hyphen-case name and a non-empty description. Run a bundled validator only if it works unmodified with already-available dependencies. Never install, vendor, or mock validation dependencies, and never claim validator success unless it actually ran. Fix structural or validation failures before reporting completion. In the final response, mention only skills that were successfully created and installed. Include one brief sentence explaining that a skill is a reusable set of instructions that can be automatically applied to matching tasks. Briefly explain that each writing-style skill activates automatically for ordinary requests to draft, rewrite, or reply in its context. Do not suggest that I must name the skill, invoke it explicitly, or say "write in my voice". For each writing context, preview `Overall voice`, `Common patterns`, and `Do and avoid`, followed by 2–3 entirely new example messages formatted as individual Markdown block quotes. End each section with a link to its full writing-style reference.

Writing-style skill: connected apps

Source: webview/assets/home-ambient-suggestions-content-2014aee9e3b2.js, offset 15499, SHA-256 1804cf5c6acbe3474003d4a081af3ef7d16239c79467decb6e1e41990ee988cd.

Exact text from the bundle. Message id home.ambientSuggestions.learnWritingStyle.prompt.connectedApps.v14.

Translator note: Hidden initial request prepared by the Learn my writing style card when communication apps are connected. It tells Codex to report non-sensitive style discoveries instead of implementation mechanics, validate without adding dependencies, and explain that the resulting skills activate automatically. Placeholder {connectedApps} is a localized list of the user's connected app mentions

Use $skill-creator to create a personal writing-style skill for each writing context represented by {connectedApps}. Keep progress updates sparse and substantive. Before sampling, limit them to the next user-relevant milestone. Once sampling begins, use them to share one or two non-sensitive style traits identified so far and how the sample is becoming more representative. Do not narrate filesystem operations, helper scripts, dependency checks, or validation mechanics. Do not ask where to install the skills; use the default personal skills directory. Analyze a representative sample of communications authored by me in each connected app. Exclude received text, quoted replies, forwarded content, signatures, automated notifications, and boilerplate. Consider variation across recipients, situations, and time. Capture stable stylistic traits—tone, sentence structure, vocabulary, formatting, level of detail, and recurring patterns—without treating topics, names, or confidential facts as style. If there is not enough evidence, explain what is missing rather than inventing conclusions. Keep email and messaging apps separate, while combining equivalent email connectors into one email skill. For each writing context, create a concise SKILL.md and a detailed `references/writing-style.md`. Configure the skill to trigger automatically for requests to draft, rewrite, or reply in that context, while avoiding overlap with the other writing-style skills. Do not reproduce or lightly rewrite source communications in generated files or responses. Remove names, sensitive details, and recognizable scenarios. Validate without adding dependencies: confirm the required files exist and that each SKILL.md has valid frontmatter with a hyphen-case name and a non-empty description. Run a bundled validator only if it works unmodified with already-available dependencies. Never install, vendor, or mock validation dependencies, and never claim validator success unless it actually ran. Fix structural or validation failures before reporting completion. In the final response, mention only skills that were successfully created and installed. Include one brief sentence explaining that a skill is a reusable set of instructions that can be automatically applied to matching tasks. Briefly explain that each writing-style skill activates automatically for ordinary requests to draft, rewrite, or reply in its context. Do not suggest that I must name the skill, invoke it explicitly, or say "write in my voice". For each writing context, preview `Overall voice`, `Common patterns`, and `Do and avoid`, followed by 2–3 entirely new example messages formatted as individual Markdown block quotes. End each section with a link to its full writing-style reference.

Write-like-me skill usage

Source: webview/assets/home-9859b126064a.js, offset 123470, SHA-256 a2c0eb8bf7416ac9470101d649035405a469efe11033acf6108e34dd34379e92.

Exact text from the bundle.

Use the write-like-me skill to find relevant examples of the user's own writing in their connected apps. Ground the draft in the selected email thread, messaging conversation, or documents. Show the draft in chat; do not send, post, or modify anything. Don't invent facts, decisions, or commitments. If there isn't enough usable context, ask one focused question.

Write-like-me sample: email

Source: webview/assets/home-9859b126064a.js, offset 97284, SHA-256 0bb195ba64ab21d48c232345a7f9f94efd4e21ac4c18699a2760077ddcd038c1.

Exact text from the bundle. Message id tpp.write_like_me.sample_prompt.email.

Translator note: Sample request displayed in the Email tab of ChatGPT Work's Writing style settings modal. Continue in chat submits this exact request using normal installed-plugin delivery. Ask for a draft based on a recent connected email conversation, not for an email to be sent. Plain text, no placeholders.

Find a recent email or thread and draft a response using my writing style.

Write-like-me sample: messaging

Source: webview/assets/home-9859b126064a.js, offset 97745, SHA-256 fa25d9061ea7b659bc69e6b91135fb2b0eea3f3600a02ee07651ecf14151c444.

Exact text from the bundle. Message id tpp.write_like_me.sample_prompt.messaging.

Translator note: Sample request displayed in the Messaging tab of ChatGPT Work's Writing style settings modal. Continue in chat submits this exact request using normal installed-plugin delivery. Message and thread mean a workplace chat conversation in a connected app. Ask for a draft, not for a message to be posted. Plain text, no placeholders.

Find a recent message or thread and draft a response using my writing style.

Write-like-me sample: documents

Source: webview/assets/home-9859b126064a.js, offset 98242, SHA-256 74809f4cd2983cd171e5218b714c4ba8d81bea8992654c7c9cf71684efd721d9.

Exact text from the bundle. Message id tpp.write_like_me.sample_prompt.documents.

Translator note: Sample request displayed in the Documents tab of ChatGPT Work's Writing style settings modal. Continue in chat submits this exact request using normal installed-plugin delivery. The source is a recent document the user wrote in a connected app; the result is a new project note based on that document. Ask for a draft, not publication. Plain text, no placeholders.

Find a recent document I wrote and draft a new project note based on it, using my writing style.

Browser extension samples

Side Chat sample: email reply

Source: webview/assets/onboarding-bd0f6da4a283.js, offset 3084, SHA-256 fc6808d0a1ede37f8c4eea117854deaa47730e2128089c43593a6c42d75516f4.

Exact text from the bundle. Message id chatgpt.work.chrome.installed.sample_work.email.prompt.sender_name.

Translator note: Prompt prefilled in Side Chat after selecting the email sample in browser extension setup. The senderName placeholder is the sample sender's display name, Samuel, and must not be translated. This page means the sample email shown in the current tab. The user must send the prompt to start the demo.

Reply to {senderName}'s message on this page

Side Chat sample: report

Source: webview/assets/onboarding-bd0f6da4a283.js, offset 4352, SHA-256 beec832532f316788ec57b4a9f925c6820c214aa8d1c2de5e006a472f1b09b09.

Exact text from the bundle. Message id chatgpt.work.extension.installed.sample_work.report.prompt.

Translator note: Prompt prefilled in Side Chat after choosing the report sample on the ChatGPT browser extension setup page. This report refers to the fictional customer research report displayed in the current browser tab. The user must send the prompt to start the demo.

Summarize this report

Side Chat sample: inventory

Source: webview/assets/onboarding-bd0f6da4a283.js, offset 5561, SHA-256 ceeb81cd1c48b64bf99f3b60a7082057cefd0285fbab89f9c6111ec94aa598a6.

Exact text from the bundle. Message id chatgpt.work.extension.installed.sample_work.inventory.prompt.

Translator note: Prompt prefilled in Side Chat after choosing the inventory sample on the ChatGPT browser extension setup page. Asks ChatGPT to mark the three existing fictional products as sold out, add five desk mousepads, and save the sample inventory. Preserve the quantities. The user must send the prompt to start the demo.

I sold out of these 3 products today but ordered 5 new desk mousepads. Update the stock and save the inventory.