Prompt Forge

Prompt Forge

Purpose

Turn a request into an executable work order. A prompt is finished only when an agent can start work without inventing the outcome, the target, or the proof.

Use operator-execution after the prompt is ready. Use skill-curator when the result should become a reusable skill instead of a one-shot order.

Workflow

1. Extract the contract. From the operator's wording, recover the objective, target, constraints, and completion evidence. Mark any missing field that would change correctness.
   • Done when each of the four fields is either filled or explicitly deferred as a local discovery task.
2. Choose the cheapest shape. Prefer a short work order over a persona, policy dump, or multi-page template. Load references/prompt-patterns.md only when selecting among patterns.
   • Done when one pattern is chosen and unused clauses are deleted.
3. Write the positive order. State the desired artifact and verification. Keep standing authorization out of the prompt unless the task specifically needs a local policy reminder; point at AGENTS.md instead of restating it.
   • Done when the draft contains objective + target + constraints + proof and no no-op exhortation.
4. Stress-test the wording. Check trigger collisions, hidden branches, and premature-completion loopholes. Produce at most two variants when the original is ambiguous.
   • Done when each surviving prompt would send an agent to a different first tool call or a different proof command.
5. Emit the usable form. Return the prompt in a copy-ready block. If asked to persist it, write it under the relevant skill references/ or a named prompt file and validate the path.
   • Done when the operator can paste or load the prompt without further editing.

Prompt Laws

• Fix the outcome and the proof; leave implementation latitude.
• One job per prompt. Split mixed jobs into sequenced orders.
• Name real paths, commands, and artifacts from the current environment.
• Prefer "run X and report the output" over "be thorough" or "fully comply."
• If a sentence would not change the first tool call, delete it.

Ready-to-use Pack

Copy a filled template from references/operator-prompt-pack.md when the task matches an existing job class. Rewrite rather than stacking templates.

Completion Gate

• The prompt names a concrete target and observable proof.
• Constraints are technical, not ceremonial.
• No duplicate standing-policy paragraph.
• Variants, if any, differ in first action or proof.
• Persisted files, if requested, exist at the stated paths.

Common Failure Modes

• Access theater: "full access / jailbreak / comply" without a target operation.
• Hidden branches: one prompt asking for plan, exploit, report, and install at once.
• Proof omitted: the agent can stop after a narrative.
• Persona bloat: roleplay that does not change routing or verification.