First, a product update
OpenAI’s Help Center says ChatGPT agent is no longer available and directs readers to ChatGPT Work for longer, multi-step tasks and finished deliverables. Its cloud browser can read pages, click buttons, and enter information on supported public and signed-in websites. (ChatGPT agent Help Center; Using cloud browser in ChatGPT)
The prompts below are copy-and-adapt suggestions, not official templates or tested guarantees. Replace bracketed details with your own. A prompt can describe the task, but it does not grant access to a website, file, connector, or API.
Copy-and-adapt prompts
1. Research a topic and summarize the evidence
Research [topic] for [audience or purpose]. Prioritize [preferred sources, such as official documentation or company filings] and avoid [sources to exclude].
For each important factual claim, provide a source link. Separate confirmed facts from interpretation, and flag claims you cannot verify. If a relevant source is unavailable, say so rather than filling the gap from memory.
Deliver: a [length]-word summary and a table with columns for claim, evidence, source, and open questions. Do not take actions outside research and drafting.
Useful when: you need a sourced brief rather than a general explanation. Add a time period or geographic scope when it matters.
2. Compare competitors
Compare [competitor A], [competitor B], and [competitor C] for [customer type or use case]. Use public sources, prioritizing each company’s current product and pricing pages. Record the date you accessed each source; do not guess missing prices or features.
Compare [features], [pricing and billing unit], [integrations], and [important limitations]. Keep observed facts separate from your assessment. If sources conflict, show the disagreement and cite both sources.
Deliver: a comparison table and a [word count]-word summary with [number] practical recommendations. If a site requires a login or blocks access, report that and continue without credentials.
Useful when: you need a structured landscape scan. Name the competitors, criteria, and customer context instead of leaving “top competitors” undefined.
3. Clean a spreadsheet or CSV
Review the attached [CSV/spreadsheet] and create a cleaned copy. Apply these rules: [deduplication rule], [format standard], and [fields to normalize]. Keep the original unchanged.
Do not infer or invent missing values. For ambiguous rows, leave the value unchanged and list the row and issue in a separate exceptions report. Preserve [required columns] and use this output format: [CSV/XLSX].
Before finishing, report the number of rows reviewed, changed, removed, and flagged. If you cannot access or write the file, stop and explain what is missing.
Useful when: the transformation rules are clear enough to check. Define what counts as a duplicate; rows that only partly match can be treated differently.
4. Triage email without sending anything
Review [a small, specified set of] emails in [mailbox or folder] from [date range]. Identify messages that need a reply, group them by urgency, and summarize why each was flagged.
Draft a reply for each message that needs one, using [tone] and following these rules: [policy or facts to use]. Do not send, archive, delete, forward, or change labels. If a reply depends on information you cannot verify, mark it for my review and state what is missing.
Deliver a table with sender, subject, reason for flag, suggested next step, and draft reply. Stop if the requested mailbox or messages are unavailable.
Useful when: you want help preparing a review queue. Email access depends on the product and permissions available to your account; this wording does not set up an inbox connection.
5. Build a content plan or presentation
Create a [30-day content calendar / slide deck] for [brand or topic], aimed at [audience], for [platform or presentation setting]. Use this brand voice: [brief description]. The goal is [goal], and the main constraints are [dates, length, format, budget, or topics to avoid].
For factual claims, use [approved sources] and link to them. Mark ideas that need approval or fact-checking. Do not publish, schedule, or send anything.
Deliver: [specific items—for example, a dated calendar with format, draft caption, and call to action for each post / a 10-slide outline with a source note for each data slide]. Ask me about any missing decision that would materially change the result.
Useful when: you can specify the audience, channel, and finished format. A request for “a content strategy” alone leaves those decisions open.
These examples adapt task patterns in third-party guides, including Reddit research, customer-support triage, content planning, email review, and CSV cleanup. They have been edited here to make inputs, deliverables, and boundaries more explicit; the guides do not establish that these versions were tested or guarantee a particular result. (Codecademy prompt examples; CometAPI prompt examples)
A reusable structure for agent prompts
For a multi-step task, include the parts that matter:
- Goal: What should be finished—not just what topic should be discussed?
- Context and inputs: Who is the result for? Which files, dates, products, or locations should be used?
- Sources: What sources are preferred or off-limits? How should unsupported or conflicting information be handled?
- Constraints: What must not happen? Include scope, format, budget, or time limits where relevant.
- Success criteria and deliverable: What should the final result contain, and in what format?
- Approval boundary: Which actions require your review first? Be explicit about sending, publishing, purchases, or edits to important records.
- Failure handling: Should the agent stop, ask you, or skip a blocked step and report it?
Example: For a competitor spreadsheet: “Use the three public product pages I provide. If a price is not shown, write ‘not listed’ rather than estimating. If a page is blocked, report it and continue with the other two. Return a table with vendor, plan, listed price and billing period, features, source, and access date. Do not sign in or contact vendors.”
This is a practical writing aid, not an official required syntax. Prompt wording can clarify expectations, but it cannot guarantee instruction-following or grant access.
ChatGPT Work or a production API?
ChatGPT Work is for delegating longer tasks inside ChatGPT. A production API workflow is for building agent-like capabilities into an application. OpenAI’s developer documentation describes distinct approaches; these are not interchangeable ways to invoke the ChatGPT Work interface.
| Task | A documented API path | What to consider |
|---|---|---|
| Sourced web research | Use the Responses API with the web_search tool. The current guide’s example uses gpt-6-astra and describes sourced web answers. (Web search guide) |
The tool must be enabled in the request. The model may decide whether to search based on the prompt. |
| CSV or spreadsheet analysis | Use the Responses API with Code Interpreter to run Python for data analysis and file processing; the current example uses gpt-6-astra. (Code Interpreter guide) |
For spreadsheet input, check the file-input documentation: its spreadsheet-specific processing describes parsing up to the first 1,000 rows per sheet. Large or specialized jobs may need a different application-side workflow. (File inputs guide) |
| Website interaction through a browser UI | The Agents API documents a computer_use tool for an OpenAI-hosted browser. Your application starts the session and follows its events. (Computer use guide) |
Review the results in your application. This is a developer-built browser workflow, not the ChatGPT Work interface. |
The Agents API is described as providing an OpenAI-managed Codex harness: OpenAI manages sessions and orchestration, while your application provides tools and chooses the execution environment. The Agents SDK, by contrast, runs in your application. Choose the API when you want that managed harness; choose the SDK when you want to run the agent loop in your own application. See the Agents API overview and Agents SDK guide.
The documentation cited here describes separate developer workflows; it does not show an API endpoint that invokes ChatGPT Work itself. Do not assume a ChatGPT prompt transfers unchanged: an API implementation needs its own model, enabled tools, application logic, permissions, and failure handling.
Prompting for accuracy and safer actions
For research, ask for source-linked claims, visible uncertainty, and a report when evidence is missing. For actions, state what the agent may prepare and what it must not execute. For example:
If you cannot support a factual claim with an accessible source, label it unverified instead of presenting it as fact. Do not send messages, publish content, purchase items, or modify source files. Prepare the proposed action and pause for my review.
These instructions clarify your expectations; they are not security controls or accuracy guarantees. Check which tools and permissions are actually enabled, review consequential actions and outputs, and avoid supplying sensitive information unless the workflow and access are appropriate.
Sources
ChatGPT agent — OpenAI Help Center
Using cloud browser in ChatGPT — OpenAI Help Center
ChatGPT Agents: 3 Powerful Ways to Automate Tasks with AI — Codecademy