Write precise instructions

Outcome and preparation

Day 1 · 10:45–12:00 · 75 minutes. Save three reusable prompts. You need the task card, source register and CSVs. Allocate 10 minutes to demonstration, 50 to writing/testing and 15 to peer review.

A good brief names role, context, task, format and checks. Request observable calculations, citations and assumptions; a long explanation alone is not proof that an answer is correct.

Step by step

1. Step 1

   Start a fresh chat and try “Write a report about Northstar.” Save the result as a baseline. Underline any claim that has no supplied source. Do not reuse invented facts in later work.

2. Step 2

   Attach data/business-kpis.csv, data/service-health.csv, data/incidents.csv and reference/kpi-definitions.md. Use the analysis prompt.

   Prompt / exampleCopy

   Role: analyst preparing a CEO/CTO weekly brief.
   Context: attached Northstar files are fictional; compare 2026-09-28 with 2026-10-05.
   Task: inventory the source rows, calculate changes and identify technical target breaches.
   Format: a table of metric, prior, current, change, formula and exact file/row/column.
   Checks: calculate with a code or spreadsheet tool; show the formula and inputs.
   Distinguish percent from percentage points. If a denominator is zero, say not computable.
   Flag missing or conflicting information. Do not infer causes or take external actions.
   

3. Step 3

   Check three numbers with your own spreadsheet/calculator: revenue growth (110000-100000)/100000*100; current error rate 3600/1200000*100; availability change 99.80-99.95. Correct the prompt if the output omits units or locators.

4. Step 4

   Use the reporting prompt on the checked analysis. Save both prompt and result.

   Prompt / exampleCopy

   Turn the checked analysis below into a one-page draft for Northstar's CEO and CTO.
   Decision: whether to investigate a four-week reliability and support pilot.
   Sections: headline, evidence, options/trade-offs, proposed actions, unknowns, review checklist.
   Cite each factual claim with its file and row/column or incident ID.
   Label recommendations as proposals. Owners and dates are to confirm, not invented commitments.
   Do not turn a correlation into a cause. No distribution or system changes.
   [Paste verified analysis]
   

5. Step 5

   Open another chat and use a reviewer prompt with the draft AND original files, not only the draft. Ask it to find errors, then personally check its findings.

   Prompt / exampleCopy

   Review this report against the attached source files as a sceptical executive reviewer.
   Return: claim, source check, issue severity, correction and evidence still needed.
   Check arithmetic, dates, units, denominators, causal claims and invented commitments.
   A report can be well written and still wrong. Do not approve unsupported claims.
   End with ready-for-human-review or needs-rework and the reasons.
   [Paste draft]
   

6. Step 6

   Save all three prompts in handover/prompts.md. Replace the specific dates and decision with clear placeholders. Run the analysis prompt again with a missing service-health file and verify that it asks for the missing evidence rather than inventing health metrics.

Expected output and verification

Three named prompts—analysis, report and review—plus before/after outputs and corrections. Expected checks: revenue +10%; current error rate 0.30%; availability -0.15 percentage points. A revised prompt should make the output easier to verify, not merely longer.

Have a partner use one prompt without your verbal explanation. If they cannot identify the required attachments, put the filenames or source placeholders directly in the prompt.

Troubleshooting

• The model fabricates citations: require file plus row key plus column, and open each locator yourself.
• A review says everything is fine: give it a deliberately wrong figure and see whether it catches it.
• Prompt is too long: separate analysis, drafting and review, preserving their input/output contract.
• Tool unavailable: perform the formula in your spreadsheet; label the AI result unverified until checked.

Completion checklist

• Three tested prompts saved with placeholders.
• Three calculations checked independently.
• Missing-source test produces an explicit gap.