SKILL.md
Plan
Shape only what is missing from the authorized intent. Prefer the caller's tracker, if any; otherwise use the caller's conversation or supplied text. A clear trivial change needs no planning worksheet or separate Plan dispatch. Planning produces no AgentOps packet.
Workflow
- Read the actual intent, relevant source owners and active constraints. Find
the caller-visible outcome, allowed write scope and first useful check. Inspect only enough context to resolve a consequential uncertainty.
- Where needed, clarify acceptance examples, important non-goals and scope in
the existing source. Include generated companions as a class: hand-edited sources plus every output of the owning regeneration commands. Include tests and live consumers that must change with them. Scope is authority, not a prediction of an exact file count.
- Choose the smallest acceptance-advancing action or discriminating check.
Identify evidence the change could invalidate and include recapture where required. Use ao provenance evidence-orphans for affected bound evidence; avoid a mandatory ledger or taxonomy for changes that do not need one.
- Revise the approach when evidence disproves an assumption under unchanged
accepted outcome and scope. Record the disproved assumption, evidence and revised check briefly in the existing source or handoff. No new permission is needed for this approach revision. Changing acceptance or expanding scope needs caller authority; never quietly weaken the original check.
- When another context needs the intent, pass enough exact source and references
to act without the author's private reasoning. The runtime binds accepted intent for final validation; useful approach notes are separate from frozen acceptance so revising a hypothesis does not fabricate acceptance drift.
A plan is sufficient when the implementer can act and the validator can judge. Then stop planning and implement. Specialists and [ground-truth routing](references/ground-truth-routing.md) are optional tools for consequential integration or design uncertainty, not universal worksheets. [Memory recall](../memory/references/recall.md) is useful only when applicable prior experience may change this work's next action.
Identity and scope
Use the runtime's source reference and digest for exact accepted intent. For conversation-only intent, existing ao provenance snapshot-intent --source - --evidence-root <explicit-root> stores resolved bytes in a caller-selected protected external non-Git evidence directory. Missing routing does not authorize a workspace fallback or a second plan artifact. Preserve legacy proof.
Scope patterns are normalized repository-relative paths, cover the behavior, and include generator-owned companions without granting unrelated directories. A live consumer outside accepted scope needs a concise exact-file amendment to the caller; continue independent authorized work while that decision is pending. [Boundaries](../rpi/references/boundaries.md) keep work/status in the caller's tracker and Git/delivery under repository policy.
Prompt
Use Plan to resolve the uncertain parser interface for this accepted change.
Keep acceptance and scope; use a real consumer check to test the assumption.
Revise the approach if it fails, then implement. No new planning artifact.
It's working if
A clear small change skips planning paperwork. A falsified assumption changes the approach and the next check; it does not trigger another approval round unless outcome or scope changes. Generated outputs remain in scope and the existing intent gives a fresh implementer enough information to act.