GocciaScript: Skill for Claude Code

Run retrospective

Assess the completed workstream through delivery-speed, process, and codebase-health lenses and agree evidence-backed improvements with the user. The actual grilling skill owns the decision loop. Present the retrospective in the conversation.

Authorization

The request to run a retrospective authorizes analysis only. File edits and other mutations require the user's final selection through grilling, and that selection authorizes only the selected documentation edits, the selected ticket actions through create-issue, and invocation of any explicitly selected immediate implementation through its normal /deliver workflow, never direct implementation by this skill. It does not authorize unrelated commits, pushes, PRs, issues, or file edits. The original request to run a retrospective is not this final selection.

Gates

• Define the workstream boundary from the current conversation, handoff, diffs, commits, issues, PRs, reviews, checks, outcomes, and rework. Record unavailable evidence and lower confidence; do not invent a narrative or broaden into a repository audit.
• Before proposing a process or design correction, trace its originating decision, the current implementation or executable configuration, and the current documentation. Distinguish missing design from implementation drift and documentation drift. Preserve a settled choice unless contradictory current evidence requires a new decision; do not re-grill it merely because the implementation or prose drifted.
• Inspect the complete available coordinator and subagent record, including process events and handoffs. Identify repeated reasoning, rebuilt or dropped evidence, loaded-but-unused or bypassed skills, manual resumes, capability-routing mistakes, and other context consumption that added no value. Report this no-value context separately from other findings and name any unavailable conversation or process evidence.
• Read references/timing-analysis.md before building the timing ledger. For a completed Milestone Rush, also read its ignored .agent/milestone-rush-events.jsonl under that reference. Reconcile telemetry with current repository, issue, pull-request, and CI evidence; partial or missing telemetry is a confidence limitation, not permission to infer values.
• Invoke grilling with the evidence and candidates. Do not imitate it with ad-hoc questions; stop if it is unavailable. Let it ask one decision at a time with a recommendation. Act only after it reaches shared understanding and confirms the exact action set.
• Assess all three lenses, even when one produces no durable finding:
  • Delivery speed: less waiting, rework, handoff friction, unnecessary scope, or cognitive load without weakening quality.
  • Process: planning, decisions, handoffs, gates, tools, and collaboration.
  • Codebase: architecture, maintainability, tests, developer experience, reliability, and accumulated friction.
• Promote only generalized, project-level lessons supported by evidence. Exclude chronology, one-off mistakes, personal preferences, duplicates, existing rules, and speculation.
• Apply de-duplication to the workstream: reuse prior investigations and decisions, coalesce the same event from multiple evidence systems, and combine candidate lessons with the same cause and action while retaining provenance. Report repeated implementation or discoverability evidence when it caused workstream friction, but do not turn the retro into a repository-wide duplication or discoverability audit; that belongs to codebase-audit.