Reasoning, Long Tasks and Breaking Problems Down

Give room to think

For analysis, maths, planning and debugging, results improve when the model works through the problem before answering. Current models (Claude's adaptive thinking, for example) can reason internally; in apps you control this with API settings such as effort. In a chat interface you can simply ask:

Work through the problem step by step, then give the final answer on the last
line as "Answer: ...".

Break big tasks into steps (prompt chaining)

One giant prompt that researches, outlines, writes and edits is hard to control. Chain smaller prompts where each output feeds the next:

1. Extract the key facts from the source documents.
2. Build an outline for the audience.
3. Write each section from the outline and facts.
4. Review the draft against a checklist and fix problems.

Each step is easier to check, and you can rerun one step without redoing everything.

Ask for verification

After drafting, check your answer against these criteria and fix any issue:
- Every number appears in the source data.
- No section exceeds 120 words.
- The call to action links to /pricing.
Then output only the final version.

Self-checks catch many errors cheaply. For anything important, add an independent check (a second prompt, a test, or a human).

Grounding and citations

Answer using only the documents provided. After each claim, add the document
title in brackets. If the documents do not contain the answer, say "I could
not find this in the provided documents."

Working with code

• Give the error message, the relevant code and what you expected.
• Ask for the smallest change that fixes the issue and an explanation of the cause.
• Run the code; models can be confidently wrong. Tests are the real judge.

Exercise

Design a three-prompt chain that turns a raw customer interview transcript into (1) a list of pain points with quotes, (2) prioritised feature ideas, and (3) a one-page summary for a non-technical founder. Write all three prompts.

Check yourself

• Why does chaining prompts make complex work easier to control?
• What does a good self-check instruction contain?
• How do you reduce invented facts in research tasks?