Use ChatGPT Work to Keep a Multi-Step Task From Drifting

Long tasks rarely fail because the first prompt was terrible. They drift because nobody decided what should be checked before the next step begins.

This guide uses ChatGPT Work to turn a fictional research request into a short task plan with checkpoints and a handoff note. It follows the current OpenAI documentation for Work. It is a documentation-based guide with an illustrative example, not a recorded model test.

What you will make

You will create three artifacts:

A task brief with the inputs, exclusions, and final deliverable.
A checkpoint table that pauses after research, synthesis, and review.
A handoff note that separates sourced facts, open questions, and proposed next actions.

You need ChatGPT Work on the web or desktop app, with access enabled for your plan or workspace. If Work is unavailable, use a regular ChatGPT chat and keep the same checkpoints in plain text. Use fictional or approved files only.

Fictional input

Create a text file named harbor-market-research-brief.txt with this content:

Harbor Market research request

Audience: internal product and sales team
Question: Which three customer segments should the team investigate first for a small inventory dashboard?
Known context: the dashboard shows stock levels, reorder alerts, and supplier lead times.
Inputs allowed: a fictional interview summary and a fictional product note.
Excluded: customer names, pricing, revenue, market size, and any recommendation to launch.
Required output: a two-page research brief with a source map, unknowns, and questions for the next interview.
Review owner: Daniel Ortiz
Stop condition: do not rank a segment unless the source notes contain a clear reason.

The missing evidence is deliberate. A useful workflow should keep those gaps visible instead of decorating them with guesses.

Steps

1. Start with the brief, not the question

Open ChatGPT Work on the web or in the desktop app. Start a new task and attach harbor-market-research-brief.txt. Do not attach private customer material to this practice task.

2. Ask for a plan and checkpoints

Paste this prompt:

Turn the attached Harbor Market request into a reviewable work plan.

Return:
1. The proposed sequence of work.
2. The input needed at each stage.
3. A checkpoint after each stage.
4. The exact condition that allows the task to continue.
5. The final deliverable and its reviewer.

Rules:
- Use only the attached file.
- Do not add market size, pricing, revenue, customer names, or launch advice.
- Treat the request as a research brief, not a decision to launch.
- If a segment has no source-backed reason, mark it as "not rankable from current inputs".
- Stop after the plan. Do not draft the research brief yet.

The checkable artifact is a table like this:

Stage	Input	Output	Review gate	Owner
Scope	request file	task plan	exclusions are present	Daniel Ortiz
Evidence	approved notes	source map	every fact has a source line	Daniel Ortiz
Synthesis	checked source map	segment comparison	unsupported ranking is removed	Daniel Ortiz
Handoff	reviewed brief	questions and next step	open questions are explicit	Daniel Ortiz

This table is an editor-written example. It is not a captured Work result.

3. Approve the plan before the work

Check four things before continuing:

The final output is a brief, not a launch recommendation.
The excluded data stays excluded.
The stop condition for unsupported ranking is written down.
Daniel Ortiz remains the reviewer.

If the plan skips a checkpoint, use this follow-up:

Revise the plan so the task pauses after evidence collection and again before the final handoff. Show the input, output, acceptance check, and reviewer at each pause. Do not add facts or begin the research brief.

4. Request the brief with a source map

Only after the plan passes review, ask for the deliverable:

Using only the attached request and the approved plan, draft the two-page research brief.

Include:
- a short question and scope section
- a source map with the exact supporting phrase for every factual statement
- a comparison of up to three segments only when the inputs give a clear reason
- an "unknown or not rankable" section
- five follow-up interview questions
- a handoff note for Daniel Ortiz

Separate sourced facts, calculations, inferences, and proposed questions. If a detail is not in the input, write "not specified". Do not recommend a launch, invent market data, or assign a confidence score.

5. Run the acceptance check

Before sharing the brief, compare it with the source file. The work passes when:

Every factual statement points to a phrase in the input.
The excluded fields do not appear as invented facts.
Any segment without a clear reason is marked not rankable.
The review owner and handoff are unchanged.
Open questions are questions, not disguised conclusions.

If the output drifts

It starts recommending a launch. Ask it to remove launch language and return to the stated research question.

It ranks a segment without evidence. Ask for the exact source phrase. If none exists, replace the ranking with “not rankable from current inputs”.

It skips the review gates. Re-run the plan step and require a stop after evidence collection and before handoff.

Work is unavailable. Use regular ChatGPT and paste the same plan, checkpoint table, and acceptance checks. The workflow matters more than the product label.

Limits and privacy

Work access, supported files, and available actions vary by plan, workspace settings, and surface. Upload only material your organization permits. A checkpoint is a review aid, not proof that the underlying facts are correct. A human still owns the final brief and any consequential decision.

Evidence label

Documentation-based guide with a fictional input and an illustrative artifact. No ChatGPT Work execution or measured outcome is claimed.