Workflow Audit

<skill_contract> <input>Bounded manual notes, prior-session evidence, or interview answers describing work, repetition, decisions, and constraints.</input> <output>A deterministic audit with mode, evidence-bounded candidates, skill/automation/loop decisions, permissions, and a typed handoff.</output> <done>Every recommendation points to supplied evidence, remains within the input bound, and has an explicit supervised promotion decision.</done> <non_goals>Editing skills, changing schedules, mutating Vault state, inventing metrics, or executing an unapproved external action.</non_goals>

Usage Template

Provide: objective, evidence source, mode (manual, session, or interview), size bound, candidate limit, operator/approval owner, and desired handoff.

Workflow

<intake> Classify the source as manual notes, prior-session history, or an operator interview. Preserve its hash and provenance. Infer a mode only when the text contains a clear session or interview signal; otherwise use `manual`. </intake>

<unknowns_gate> Separate observed repetition from a proposed improvement. Missing frequency, owner, verifier, permission, or stop condition makes a candidate review-only; it does not justify a schedule or autonomous loop. </unknowns_gate>

<execute> Extract a bounded list of distinct work candidates. For each candidate, decide:

• SKILL_CANDIDATE when the work has a reusable procedure or multiple steps; otherwise KEEP_MANUAL.
• AUTOMATION_CANDIDATE only when repetition or a clear trigger is evidenced; otherwise HOST_REQUIRED_REVIEW.
• LOOP_CANDIDATE only when retry, monitoring, polling, follow-up, or a repeat-until condition is evidenced; otherwise NO_LOOP.

Record the evidence snippet, approval owner, verifier, stop condition, and side-effect expectation. Promotion is a separate supervised action with a receipt and rollback path. </execute>

<evaluate> Reject recommendations that rely on an unbounded input, an inferred external permission, a fluent success claim, or a video/demo as authority. Check candidate count, source hash, evidence linkage, and `side_effect_count: 0`. </evaluate>

Failure Protocol

Return NEEDS_INPUT for missing evidence or a required authority decision, INSUFFICIENT_EVIDENCE when repetition cannot be supported, and VERIFY_FAILED when a recommendation cannot be traced to the supplied source. Preserve the audit receipt and do not promote.

Output Contract

Return status, result, evidence, unknowns, and next_action. The result includes mode, source hash, bounded candidates, skill/automation/loop decisions, approval, verifier, stop condition, and side-effect count.

Edge Cases

• A quiet prior session is not proof that a workflow is eligible for automation; record INSUFFICIENT_EVIDENCE or request a frequency probe.
• A recurring task that writes to a shared Vault remains a supervised staged-write candidate even when the extraction itself is deterministic.

Success Metrics

• Source provenance and candidate bounds are preserved.
• Every decision is evidence-linked and independently checkable.
• No skill, schedule, external service, or Vault file is changed by the audit.

Quality Gates

• Mode is explicit or conservatively inferred.
• Unknown permissions, verifiers, and stop conditions are surfaced.
• Promotion is supervised and receipt-backed.
• side_effect_count is zero.

</skill_contract>