Practical knowledge. Smarter decisions.AR  ·  EN  ·  FR

Writing better AI prompts for practical work

A repeatable prompt framework with context, output requirements and a verification checklist.

AI output improves when you specify the task, input boundaries, audience and success criteria. Never confuse a fluent answer with a verified one.

Computer with programming code and contextual menu on screen
Real editorial photograph — Daniil Komov / Unsplash. Illustrative photograph, not live market data or a product endorsement.

Give a precise task and context

Instead of “analyze sales,” provide the period, currency, column definitions and question. Tell the model not to invent missing data. Review confidentiality policies before sharing private documents.

Ask it to list assumptions and unresolved questions rather than guessing.

Specify the format

Request an exact table structure, checklist or draft. Include one short example if needed. Longer output is not inherently more accurate.

Language models can make confident mistakes.

  • Objective and audience
  • Constraints and reference material
  • Output format and verification rule

Verify consequential claims

Use a separate review pass to flag unsupported numbers and omissions. Recalculate arithmetic independently and confirm legal or financial claims against dated primary sources.

AI assists professional judgement; it should not replace it.

Reusable template

Try: “Act as a [role]. Task: [specific outcome]. Inputs: [known information]. Constraints: [limits]. Output: [format]. Identify assumptions and ask before inventing missing facts.”

Iterate by changing one unclear requirement at a time.

Worked example: summarize sales without inventing records

Imagine a synthetic spreadsheet with exactly three columns: invoice date, sales amount and city. Instead of asking the model to analyze the file vaguely, require monthly and city totals, a count of rows rejected for missing dates, and a separate exception report. Explicitly prohibit filling blank values. Specify the currency, output columns and the control that the sum of city subtotals must reconcile to the grand total. This is a teaching scenario, not a claimed result from a real customer's records.

Recalculate the totals independently in Excel or Python. Prepare three small test cases: a valid row with a positive amount, a row missing a required value, and a duplicated invoice. Define whether duplicates should be rejected before evaluating the assistant. Check both the numerical answer and the handling of exceptions; polished wording is not an accuracy test.

Protect confidential inputs and version the prompt

Before entering work data into any AI tool, identify customer names, account numbers, personal identifiers and confidential commercial terms. Begin with representative synthetic records that preserve the table structure but not the identities. If real sensitive information is essential, confirm the organization's permission, contractual protections and data-retention settings; do not assume a general chat session is an approved financial-document repository.

Keep a versioned record of the instruction, allowed inputs, expected output, evaluation examples and reviewer decisions. When changing one instruction, rerun the same fixed cases and compare the observed errors and omissions. Do not claim that a revised prompt performed better without a repeatable test. This creates an auditable workflow instead of relying on an unverified first impression.

Official resources

Try a practical free calculator or builder ↗
Important notice

No online income or trading system guarantees profits. Results depend on skills, costs and circumstances.

Disclaimer ↗

Search the library