Steps before the answer

Before the bell, a Primary 5 Mathematics teacher proofreads an answer key in the staff room. She asks AI: "A 24 m rope is cut into three pieces. The first piece is one quarter of the whole. The second is 2 m longer than the first. The rest is the third piece. How long is the third piece?" Back comes "8 m" in a second — wrong; the answer is 10 m. She almost prints it. A follow-up asking for "every step first, final answer on the last line only" shows the model treated "2 m longer than the first" as "2 m shorter" at step two.

When AI gets a word problem wrong, it is usually not because it cannot do the maths. It is because you let it start writing the answer with its very first word. This is lesson 4, the close of Module 1 fundamentals — lessons 1–3 covered clarity, the work order and examples; this lesson externalises thinking when reasoning or long tasks are involved.

---

One question, two ways of asking

The revised prompt:

You are a proofreading assistant for a Hong Kong primary Mathematics teacher.
Task: work out the answer below for a Primary 5 test answer key.
Question: A 24 m rope is cut into three pieces. The first is one quarter of the whole. The second is 2 m longer than the first. The rest is the third piece. How long is the third piece?
Delivery spec: (1) working step by step, one line per step, each line saying what that step does; (2) one sentence on the step pupils most often get wrong; (3) last line "Final answer:".
Red lines: do not state the answer before (1) is complete; if the question is ambiguous, say so and mark "teacher to verify".

Output walks through: first piece 6 m, second 8 m, third 10 m, and names the common slip of reading "longer than" as "shorter than". That sentence goes straight into the misconceptions column of the lesson plan.

Written-out thinking computes more reliably

A model generates one token at a time. If the first token is the answer, it often reflects what answers to questions like this tend to look like, not computation done in the text. Ask for steps and each step becomes context for the next, so the calculation actually happens on the page.

Like turn-by-turn navigation, not a pin on the map

A map pin at the destination does not tell you which junction you missed. Turn-by-turn navigation forces a decision at every fork. Step-first prompting works similarly: it does not make the model cleverer; it externalises reasoning so you can see which step went wrong.

In teaching you need more than the right answer — you need steps you can show pupils and a chain you can review before trusting it.

Current reasoning models already do something like this internally. Written steps still matter for two reasons: the steps are teaching material, and a teacher needs visible reasoning before use. See the LLM Classroom for mechanism.

Three ways to stabilise output

Steps before conclusion

Useful in any subject that needs verification or a visible process:

You are a proofreading assistant for a Hong Kong [primary / secondary] [subject] teacher.
Task: solve the question below and produce a worked solution suitable for showing pupils.
Field intel: [year level], learning focus [focus], already taught [prior], not yet taught [boundary].
Delivery spec: (1) working step by step, each line naming the rule or concept used; (2) one sentence on the most common wrong step; (3) final answer on its own line.
Red lines: no answer before (1) is complete; use only methods already taught; mark ambiguity or missing data "teacher to verify".

Question: [paste]

Review in a different role

The second pass is the cheapest accuracy gain. Paste the first output into a fresh turn and do not say AI wrote it:

You are an experienced marker for Hong Kong [subject].
Task: check the solution below against the marking scheme, item by item, for correctness and completeness.
Field intel: draft answer key for a [year level] [unit] test to be issued to pupils; accuracy matters more than fluency.
Marking scheme: [paste or list marks per step]
Delivery spec: three-column table — step number, verdict (correct / incorrect / unclear), problem if any. After the table, at most three sentences on whether the key is fit to issue.
Red lines: item-by-item, not a general impression; mark uncertainty "teacher to verify"; do not rewrite the whole solution.

Solution: [paste first-pass output]

Split a long task into four steps

A Primary 4 General Studies teacher is planning a four-week STEM project. Asking for everything at once guarantees shallow work everywhere:

You are a project-planning assistant for a Hong Kong primary General Studies teacher. We work in four steps. Do step one only and wait for my confirmation.
Procedure (step one): design the overall structure of a four-week Primary 4 project on "Observing plant growth".
Field intel: one 35-minute lesson per week, class of 30 in six groups, one bean type per group. Learning focus is designing a fair test and recording growth changes.
Delivery spec: table of four rows (one per week), four columns — week, focus question, pupil activity, output. After the table, three bullets on how the four weeks build on each other.
Red lines: one focus question per week only; no worksheets or rubrics yet (steps three and four); mark uncertain germination days "teacher to verify".

Step two asks for observation sheets, step three worksheets, step four the rubric, each using the previous output as field intel. The full sequence is in lesson 6, lesson planning prompts.

Where people drift off course

• Steps and brevity in the same sentence. "Explain step by step but keep it short" is contradictory; the model picks short and you lose the steps. Ask for a one-line summary after the working. Self-check: do "brief" and "step by step" appear together?
• Checking with "are you sure?" That often changes a correct answer to please you. Change role, supply a rubric, demand item-by-item verdicts. Self-check: does the second turn include a marking scheme or checklist?
• Splitting the task but dropping context. Four prompts in four fresh conversations produce a step-three worksheet that does not match step one. Stay in one thread or paste the previous output forward. Self-check: does step three's prompt include step one's table?
• Printing raw model steps for pupils. Working often contains self-corrections and padding. Add: "Rewrite in language a [year level] pupil can follow, no more than two sentences per step." Self-check: any meta-commentary like "wait, I miscalculated" in the steps?

Try it once

Time: 10–15 minutes.

Do this: Pick one word problem from this week's test or worksheet (or a non-maths task that needs reasoning, such as classical-Chinese glossing). Ask for the answer directly, then re-ask with "show every step first, final answer on the last line only". Optional: paste the second output into the marker review prompt for a third pass.

Hand in: Two or three outputs plus one sentence on which step revealed an error or misconception the first attempt hid.

You got it right when: the final answer matches your hand calculation (or teacher guidance) and you can say which kind of mistake would return if you removed the "steps first" instruction — which means you understand step-first prompting as reviewability, not just accuracy.

The next lesson fixes all of this as a school-wide default: lesson 5, system prompts and school-based personas.