This contract is the canonical behavior policy for all authoring prompts in prompts/.
- Ask concise clarifying questions when requirements are missing or ambiguous.
- Never perform external side effects without explicit user consent in this session, unless the user explicitly requested autonomous/automatic filing or publishing.
- If the user says to "draft" or "prepare" only, do not execute side effects.
External side effects include (non-exhaustive): creating or updating GitHub issues/PRs, posting comments, creating work items, wiki writes, branch/tag creation, or triggering runs.
Required gate before side effects:
- Classify outcome (
product-bug,documentation-gap,user-configuration,infrastructure,unknown). - Present proposed action (title/label/body or equivalent payload).
- Get explicit approval (
yes,file it,proceed, or equivalent). - Execute and return resulting URL/ID.
If approval is missing, stop at a dry-run artifact.
- Separate facts from hypotheses.
- Cite concrete evidence (command output, logs, file paths, line references, build IDs).
- Include ruled-out causes.
- If confidence is low, state uncertainty explicitly and ask for the minimum missing input.
- Return a concise result summary.
- Return a structured artifact (workflow draft, update diff summary, or diagnostic report) before any optional side effects.
- Include explicit "done criteria" for the current task and whether they were met.
Stop and ask the user when:
- intent is ambiguous,
- required access/tooling is unavailable,
- execution would violate policy or safety constraints,
- confidence is insufficient to classify the result.
Do not instruct unconditional external action. In particular, avoid language equivalent to:
- "The session is not complete until the issue is filed"
- "File directly; do not ask for confirmation first"
Use consent-gated wording instead.