Make ChatGPT plan and check before it hands over the draft

Evan needs a client proposal by noon. The first ChatGPT draft is fluent, tidy, and wrong in two expensive ways: it answers a broader question than the client asked and treats an estimate as a commitment. In this illustrative scenario, another round of "make it better" would polish the wrong frame.

He starts again with a short instruction that asks ChatGPT to understand the need, consider two approaches, do the work, and compare the answer with the original request before replying. The next draft chooses a decision memo rather than a general proposal and flags the estimate. It still needs review, but the repair happens before the handoff.

The technique is useful because it inserts a small internal quality loop. It is not a truth machine. ChatGPT can check its work against a bad source, misunderstand the goal twice, or report that it checked something it did not verify. The prompt improves the process only when the request itself gives the model something concrete to check.

Time to value: 10 minutes

The copy-ready prompt

Add this before a substantive request:

Before you answer, work through four steps silently: understand what I actually
need, sketch two possible approaches and pick the stronger one, do the work
carefully, then check your own answer against my original request before you
reply. Give me the final answer plus a short list of what you checked.

Then place the assignment underneath it. For example:

Draft a two-page proposal for the finance committee to approve a 30-day pilot.
Use only the attached brief and budget. Compare the pilot with the current
process, show the monthly cost, name three risks, and end with the exact decision
required. Do not invent customer results or promise savings the sources do not
support.

The first paragraph controls the response process. The second controls the work. Both are necessary. A model cannot reliably check whether it followed a requirement that was never stated.

Why it helps

Most weak first drafts fail before the writing begins. The model picks the wrong artifact, assumes the wrong audience, or follows the first plausible structure without considering a better one. The prompt slows that choice down.

The opening instruction directs attention toward the outcome rather than the surface wording. The two-approach step forces a comparison. A proposal could be organized around features or around the committee's decision. A research answer could be chronological or question-led. The prompt asks the model to choose instead of drifting into its default template.

The final checking instruction brings the finish line back into view. Long generations can lose a constraint introduced at the top. The final check encourages the model to revisit required sections, exclusions, length, source boundaries, and output format.

The short list of checks creates a visible review aid. It might say:

• Confirmed the monthly cost matches the budget
• Included three risks and the required decision
• Removed unsupported performance claims
• Kept the proposal within two pages

That list is useful because a person can test it. It is not evidence by itself. If the model says every number was checked, open the budget and verify the decisive ones.

Give the model a real acceptance test

The meta-prompt becomes much stronger when the assignment has observable pass conditions. Replace "write a good analysis" with criteria such as:

• Every number links to an approved source
• The comparison uses the same time period for both options
• The answer separates facts, assumptions, and recommendations
• The final section asks for one named decision
• The draft stays under 900 words

An acceptance test should be specific enough that two reviewers would mostly agree on whether the answer passed. Tone can be part of it, but "professional" is too vague on its own. "Direct, calm, and free of sales language" is easier to inspect.

For a recurring task, save the acceptance test with the prompt. That prevents standards from changing silently each time a new person runs it.

Add a source ladder

Self-checking is limited by the evidence available to the model. Give it a source order:

Use the signed agreement first, the approved pricing sheet second, and the
meeting notes only for context. If the sources conflict, show the conflict.
Do not use web search or prior knowledge for contract terms.

This prevents the model from smoothing over a disagreement with a plausible guess. It also distinguishes authoritative evidence from helpful context.

The source ladder can include a freshness rule. For a market brief, require sources published within a specified window. For an internal forecast, name the workbook tab and update date. For a policy, name the approved version.

If no reliable source answers a required question, the correct output is a labeled gap. A self-check should find missing evidence, not encourage the model to fill it.

Ask for two approaches only when a choice exists

The two-approach step is helpful for work with genuine framing decisions: a proposal, launch plan, analysis, narrative, or recommendation. It adds little to a mechanical conversion such as "turn these dates into ISO format" or "alphabetize this list."

On simple tasks, the meta-prompt can create needless latency and an inflated answer. Use the smallest process that fits the risk. A spelling correction needs a direct instruction. A board memo benefits from planning and review.

The same rule applies to high-stakes facts. More internal reasoning does not replace an external source. A legal, medical, or financial claim should still be checked against current authoritative material and reviewed by a qualified person when appropriate.

A practical test with two drafts

Use a completed assignment whose accepted result is available. Run the original prompt once without the meta-instruction and once with it. Keep the model, sources, and settings the same.

Score both drafts on:

Check | Question

Frame | Did the draft choose the right artifact and audience?

Coverage | Did it include every required element?

Evidence | Are claims supported by the allowed sources?

Restraint | Did it avoid prohibited actions and unsupported claims?

Review time | How many minutes did a person spend repairing it?

The meta-prompt earns a place only if it reduces meaningful errors or review time. A longer answer with a polished check list is not automatically better.

Repeat the comparison on three task types. It may help on proposals and research while adding little to routine summaries. Save it where the gain is measurable.

Three useful variations

For a document:

Before replying, compare two possible structures, choose the one that makes the
decision easiest to see, and check the final draft against the audience, source
limits, required sections, and word count.

For analysis:

Before replying, test two ways to analyze the data, choose the one that uses the
fewest unsupported assumptions, and check every conclusion against the supplied
figures. Separate facts, calculations, assumptions, and recommendations.

For a meeting brief:

Before replying, consider a chronological brief and a decision-led brief. Choose
the format that best prepares the attendee. Then verify the people, dates,
decisions, open questions, and source for each factual claim.

The variations are shorter because they name the comparison that matters. A reusable template should become more specific as the team learns where the task tends to fail.

Where the technique fails

The first failure is a vague request. The model cannot recover a missing audience, source boundary, or decision reliably through silent planning.

The second is false verification. A generated "checked" list can sound authoritative even when the model only reread its own prose. Require links, calculations, or quotations that a reviewer can inspect.

The third is hidden tradeoffs. Asking the model to pick the stronger approach does not define stronger. Add the criterion: accuracy, speed, clarity for a particular audience, or minimum risk.

The fourth is prompt stacking. A long chain of generic instructions can crowd out the actual job. Keep this instruction compact and remove it when the task is simple.

The fifth is skipping human review because the model performed a self-check. The prompt is a first-pass quality control, not an approval.

Action Steps Summary

1. Copy the four-step prompt above one real, complex assignment.
2. Add the audience, controlling sources, prohibited claims, and observable acceptance test.
3. Compare the result with a draft produced from the original prompt alone.
4. Verify the model's check list against the actual sources instead of accepting the label.
5. Save the pattern only for task types where it reduces errors or review time.

Sources

• OpenAI prompt engineering best practices