Prompt patterns

Pattern 1: step by step#

Work in this order:
1. Name the facts that appear in the text.
2. Name the facts that are missing.
3. Only then: answer the question, solely from 1.

The gain is not in the thinking but in step 2 making the gaps visible before step 3 could fill them.

Pattern 2: exit clause#

If the answer does not follow from the material supplied,
answer solely with: NOT IN MATERIAL.
Never guess.

The refusal rate becomes a metric. A system refusing in eight percent of cases is more useful than one that always answers and is wrong eight percent of the time.

Pattern 3: role separation#

One call produces, a second checks, with a different framing:

Call 1: Draft an answer based on the sources.
Call 2: Check whether every statement in the draft is supported by the sources.
        List unsupported statements verbatim.

A model asked to judge its own text in the same pass usually confirms it. A separate call with a checking brief finds considerably more.

Pattern 4: tool instead of arithmetic#

Do not calculate yourself. Return every calculation as an expression
to be evaluated by a calculator:
{"compute": "1250.50 * 0.20"}

Pattern 5: few-shot examples#

Two to five examples of the input and output you want, including at least one where the correct answer is a refusal. Without that example the model learns to always answer.

Self-consistency#

Ask the same question several times at temperature above zero and take the most frequent answer. The real gain is not the answer but the level of agreement: it is an empirical confidence measure.

Cost: a factor of five to twenty. Worth it only where an error is expensive.

What does not work#

The chain of thought is not evidence#

When a model prints its reasoning, that is a working aid and not a log. The printed route need not match the actual computation. For a justification in the legal sense, the result is checked against the sources, not the model's narrative. See Reasoning models.

Pattern cards to take away#