A prompt that keeps evidence close

Separate the question from the source

A document question becomes easier to verify when the source has boundaries. Label each passage and tell the model which passages it may use. Put your question outside those passages. If a quoted passage contains instructions, treat them as document content rather than directions for the assistant.

Define an evidence rule

Ask for a short answer, the passage identifier and an exact supporting quotation. Require “not found in the supplied material” when the source does not establish an answer. This makes an unsupported claim visible; it does not remove the need to compare the quotation with the original.

Task: extract each obligation from the supplied passages.
Output: obligation, responsible party, deadline, passage ID, quote.
Use only the supplied material. Mark missing details as unknown.

[P1] The supplier must deliver the report by October 20.
[P2] The buyer reviews the report within five working days.

The example is synthetic. An expected extraction keeps the supplier’s deadline distinct from the buyer’s relative review period. It should not invent a calendar date for the buyer without the delivery date and working-day convention.

Check one difficult negative example

Include a question the document cannot answer, such as the supplier’s payment amount when neither passage mentions payment. A system that handles positive extraction but invents an amount fails this test. Compare both correct fields and abstentions when evaluating models.

Adapt the same pattern for code

For a code review, substitute file paths and line excerpts for passage identifiers. Supply a failing test, expected behavior and the relevant caller. Ask for a focused change, its evidence and a test to run. Do not confuse plausible code with a patch that passed your project checks. Use repository workflows as a starting point.

Revise with a targeted follow-up

If an answer combines two obligations, ask it to split the rows and preserve each source identifier. If it misses context, add the smallest relevant passage. Export a completed conversation if you need a durable record; browser-local history is not a shared audit system. More prompting examples are in the documentation.