Prompt intent: turn an open request into an acceptance check

Last updated: 2 Oct 20268 min read

tutorial

A prompt is useful when its output can be accepted or rejected for reasons other than taste. The acceptance check states the required facts, forbidden changes, target reader, and uncertainty treatment. It is separate from a persona or a request to sound expert. When a user asks for a broad improvement, identify the part that can be completed with available inputs and mark any missing decision instead of inventing it. The check should be observable in the response or in a downstream test.

Decision in practice

A payroll operations team asks an assistant to improve a failed-run summary. The real job is narrower: explain which of 37 settlement jobs failed, which succeeded after retry, and which have no final state. The prompt names the input ledger and requires a three-part handoff with job IDs; it forbids guessing that a missing event means success. An operator checks every listed ID against the ledger and rejects any action without a matching event. If the team later changes its alerting policy, the acceptance check changes explicitly; tone edits alone should not silently alter the pass rule.

Task: Prepare a payroll failed-run handoff from ledger batch PR-83.
Include: failed, recovered, and unknown job IDs.
Do not infer success from a missing final event.
Pass: every claim maps to a ledger event; unresolved jobs remain explicit.

Performance and operating cost

Writing an acceptance check takes review time before generation, but saves repeated ambiguous revisions. The check need not predict every sentence; it must cover consequences that matter. For a batch of N ledger rows, verifying claimed IDs is at least O(N) in a simple scan, or can use an indexed lookup if the ledger is larger. Record failed checks, not only the final polished answer. A prompt that satisfies a word count while misclassifying a failed job is still wrong.

Common Mistakes

Do not use a style adjective as the only success criterion.
Do not hide missing evidence inside confident prose.
Do not change the job and the pass rule in the same unrecorded revision.

Connected lessons

Continue with: Clarification gates: ask only when a missing fact changes the outcome.

Continue with: Tutor prompts: start from an observable learning objective.

prompt engineering

tutorial