Seven techniques
General information, not legal advice. Check your own bar or regulator's guidance and your client engagement terms before putting any client information into an AI tool.
- 1
Give an example
Paste one invented example of a good output and say 'follow this format'.
- 2
Ask for stages
'First list the issues, then list facts needed, then suggest next steps.' Stages make errors visible.
- 3
Constrain the output
'Return a table with these exact columns.' Predictable shapes are easier to paste into Notion.
- 4
Require sources from your text
'Quote the sentence you relied on.' You can then check against your own material.
- 5
Add a critique pass
'Now review your answer as a sceptical senior lawyer and list weaknesses.'
- 6
Set refusal rules
'If the text does not say, answer "not stated" rather than guessing.'
- 7
Test and version
Run the prompt on three invented cases and record changes.
A worked example
Goal: extract key dates from a non-confidential agreement. Prompt: 'You are a careful paralegal. From the text below, extract every date, deadline or time period. Return a table with Item, Date or period, Exact quote, Section. If an item is ambiguous, write "unclear". Do not calculate dates. Text: [text].'
Why it works: it gives a role, a precise task, a fixed table, a quote column for checking, a rule for uncertainty and a ban on calculation, where assistants often err.
Testing prompts
| Test | What you learn |
|---|---|
| Run on an easy invented case | Whether the basic structure works |
| Run on a tricky case with an exception | Whether it misses edge cases |
| Run twice | How consistent the output is |
| Run in another assistant | Whether the prompt depends on one model |
Store tested prompts with a version and test date in a Notion database so your team reuses what works.
Limits of technique
A beautifully engineered prompt can still produce a wrong answer. The assistant has no responsibility; you do. Keep human review, keep data anonymised, and avoid using AI for steps where one error is costly, such as limitation periods.
Frequently asked questions
- Is prompt engineering a real skill for lawyers?
- It is mostly clear writing and good instructions, which lawyers already practise. The new part is testing.
- Do I need special tools?
- No. A chat assistant and a place to store prompts are enough.
- Do these techniques work for every model?
- Generally, but results vary. Test in the assistant you use.
- Can prompting prevent fabricated citations?
- It can reduce them by asking for no citations, but the safest approach is to verify everything independently.
About the author
Raghav Handa trained as a lawyer, then spent years building legal technology, from enterprise legal products to custom systems for law firms.
Keep reading
- LibraryHow to write AI prompts as a lawyer: the five-part formulaA simple five-part formula for writing AI prompts in legal work, with before-and-after examples, common mistakes and a checklist for reviewing what comes back.3 min read · 17 October 2026
- AIAI prompts for lawyers: 14 copy-and-paste prompts with placeholdersA library of 14 AI prompts for lawyers with [placeholders]: summaries, issue lists, email drafts, checklists and review prompts. Use them on non-confidential text.4 min read · 17 October 2026
- AIAI workflows for lawyers: five repeatable, reviewable processesFive practical AI workflows for a law practice, from intake triage to weekly reviews, each with a human checkpoint, a data rule and a place to record the result.3 min read · 17 October 2026
- LibraryClaude chat for lawyers: a practical guide to getting good answersA practical guide to using Claude's chat interface in a law practice: setting up a conversation, giving context, checking output and keeping client data safe.3 min read · 17 October 2026