Agree on the result first#
For a substantial task, start with a brief stating the goal, materials and quality criteria. Ask for a plan and review it before processing all the material. This method does not require a particular model; check tool availability in your product.
I need [output] for [audience]. The materials are [list]. Propose at most five steps. For each, state the deliverable and how I can check it. First list missing information that could change the conclusion.
The result should be an executable plan. If the first step needs an inaccessible source, resolve access or adjust the scope. The plan is not the finished work.
Research with traceable claims#
- Define the question, country and time range. Separate new events from older context.
- Request primary sources and a direct link for every important claim.
- Open the sources. Compare the publication date, event date and scope.
- Ask for conflicts and gaps. A second article repeating the same press release is not independent confirmation.
- Only then produce a summary for your audience.
Research [question] for [country] during [from–to]. Distinguish confirmed facts, one party’s claims and uncertain information. For each important point, provide a direct source and the event date. Mark sources you cannot open. Do not draw conclusions from headlines alone.
Check: manually verify three claims that would most affect your decision. Citation count is not a measure of accuracy.
Write without losing your voice#
Provide an outline, a short sample paragraph and facts that must stay unchanged. Approve the structure first, then work on each part. After assembling the text, check repetition, transitions and consistent terminology.
Edit the following text for [audience]. Preserve its meaning and all numbers. Shorten sentences where this improves readability. Add no new claims. After the draft, list changes that might affect tone or meaning. Text: [paste]
If the draft feels generic, add a concrete situation or example. “Write it better” alone does not define a useful criterion.
Learn by making your own attempt#
Say what you already know and ask for questions one at a time. Do not ask for all the exercise solutions at once. Make your own attempt, then ask for an explanation of the specific mistake.
I am learning [topic]. I understand [basics] and struggle with [problem]. Explain one idea using an example. Then give me one short exercise and wait for my answer. If I make a mistake, identify the wrong step rather than only giving the correct result.
Check the explanation in a textbook or another reliable reference. For languages, verify unusual expressions in a dictionary; for maths, check substitution and units.
Data analysis and programming#
For a table, explain column meanings, units, dates and missing values first. Review data-cleaning rules before calculation. Verify totals or averages independently in a spreadsheet. An association between numbers does not establish causation.
For code, provide the language, environment version, exact error and a small reproducible example without secret keys. Request a small fix and an explanation of its effects. Run changes on test data first.
Find the cause of this error in [environment and version]. Expected behaviour: [description]. Actual behaviour: [description]. Example and error: [paste]. Suggest the smallest fix and a test that exposes the original problem. If you have not run the code, say so.
If an answer sounds convincing but lacks a calculation or test, ask for a verifiable procedure. An assistant saying it ran something is not itself an execution record.