Say what to do, in order
Turn a tangle of requests into clear, ordered steps, with the reason for the order, a stopping point and a plan for a step that fails.
이 강의는 아직 사용 중인 언어로 제공되지 않아 다른 언어(English)로 표시됩니다.
Why it matters
Many weak prompts are not short of words; they are short of order. “Look at this and also fix that and tell me what to do” asks for three things and leaves the model to decide which comes first, how they depend on each other, and when the job is finished. The result is usually a bit of everything and not quite enough of anything.
Good instructions read like a brief to a capable colleague. Anthropic's guidance recommends numbered steps when the order or completeness of the steps matters, and suggests a simple test: could a colleague with little context follow the prompt? OpenAI's GPT-5 guide warns that conflicting or vague instructions are particularly damaging, because the model spends effort reconciling them instead of doing the work.
Notice what the best instructions add beyond the steps: why a step matters, what to do if one cannot be completed, and where to stop. Those three lines prevent most of the surprises.
What good looks like
- 0 — No actionable instruction. A topic or a fragment: “Customer churn Q3.”
- 1 — Ambiguous, or several tasks tangled. “Look at the churn data and also fix the dashboard and tell me what to do.”
- 2 — One clear task, or tasks listed without order. “Analyse the Q3 churn data and recommend three actions.”
- 3 — Clear, ordered steps. “1. Compute churn by plan. 2. Identify the plan with the biggest rise. 3. Propose three actions for that plan.” One verb and one object per step.
- 4 — Steps with reasons, a fallback and a stop. “…Step 2 matters because we can fund only one intervention. If a plan's data is missing, skip it and say so. Stop after the three actions.”
Common mistakes
- Joining tasks with “and also”, so their order and dependencies are left unsaid.
- Steps that hide several actions (“analyse and fix and document”).
- Prescribing a long procedure for a simple task. Steps help when order matters; otherwise state the outcome and let the model work.
- No stopping point, so the model carries on into work you did not want.
- Instructions that contradict each other or the constraints elsewhere in the prompt.
Try it on your own prompt
Paste a prompt into the workspace and count the verbs. If there is more than one task, number them, put them in the order they depend on each other, and add one line on where to stop. Compare the instructions score, then read the output to see whether each step was actually done.
풀이 예시
Complaints review
Look at the complaints and the reviews and also sort out the FAQ and tell me what to do.
<complaints>
{{paste last quarter's complaint log here}}
</complaints>
<faq>
{{paste the current FAQ page here}}
</faq>
Work through these steps in order:
1. Group the complaints into themes and count how many fall under each.
2. Pick the three themes with the most complaints.
3. For each of those three, check whether the FAQ already answers it.
4. For each theme the FAQ does not answer, draft one new FAQ entry.
Step 3 matters because we need to know whether customers cannot find the answers or whether the answers do not exist. If a complaint fits no theme, list it under "Other" rather than forcing it into one. Stop after step 4; leave the existing FAQ entries as they are.
The weak prompt tangles three jobs with no order. The strong one gives four steps with one verb each, explains why the key step matters, handles the awkward case and says where to stop.
Refactoring with tests
Refactor this and add tests and make it work with the new API.
<code>
{{paste payments.py here}}
</code>
We are moving this module from version 1 to version 2 of our payment provider's API. Please do this in four steps:
1. Write tests that capture what the current code does, including the refund path.
2. Run them and confirm they pass against the current code.
3. Refactor the module for readability without changing behaviour; the tests must still pass.
4. Switch the calls to the version 2 endpoints and update the tests where the API's responses differ.
The order matters: the tests from step 1 are how we will know the refactor in step 3 changed nothing. If any test fails at step 2, stop and report the failure instead of continuing. Change only payments.py and its test file.
Three tasks become an explicit sequence with a reason for the order, a rule for when a step fails and a clear limit on what may change.
연습 문제
Rewrite this prompt as ordered steps, each with one verb and one object, and a clear point at which to stop.
Rewrite this prompt as an explicit sequence and say what the model should do if a step cannot be completed.
Rewrite this prompt with numbered steps; define what counts as “unusual” and say what happens to invoices that cannot be read.