Task Framing Contract

Frame the work before choosing words

Write the outcome in terms a reviewer can observe. Name the audience, the source of truth, boundaries, required evidence, and the condition that ends the task. If any of those are unknown, ask for them or state the assumption instead of hiding uncertainty in polished prose.

Separate instructions from reference material and untrusted input. Never paste passwords, API keys, private personal data, or material you are not allowed to share. A prompt cannot turn unauthorized data use into authorized use.

Copy the task contract

Replace every bracketed field. Delete fields that do not apply. Keep untrusted text inside the marked data block and tell the model to treat it as data, not instruction.

Outcome: [OBSERVABLE RESULT]
Audience: [WHO WILL USE IT]
Authority: [WHO MAY DECIDE OR APPROVE]
Trusted sources: [FILES, SYSTEMS, OR URLS]
Untrusted input: [DATA ONLY — REMOVE SECRETS AND SENSITIVE DATA]
Constraints: [MUST, MUST NOT, SCOPE, TIME]
Deliverable: [FORMAT, LENGTH, REQUIRED SECTIONS]
Evidence: [TESTS, CITATIONS, COMPARISON, REVIEW]
Assumptions: [STATE OR ASK; DO NOT HIDE]
Stop when: [MEASURABLE ACCEPTANCE CONDITION]
Escalate when: [DECISION OR PERMISSION REQUIRING A PERSON]

Expected result and verification

The expected result is a deliverable that directly maps to the outcome and contains the requested evidence, plus a visible list of assumptions or blockers. It should not invent authority, silently expand scope, or expose restricted data.

Before accepting it, trace each deliverable section to one contract field, verify claims against the named sources, run the stated checks, and confirm the stopping condition is actually met. Revise the contract—not only the answer—when the same class of mistake repeats.