agentleFS
frostneyfrostney/pascal-mcp-sdk/.agents/skills/run-retro/SKILL.md

run-retro

Skill2 starsChanged 3 months agoMITView on GitHub

Reviews a completed workstream from conversation, repository, and forge evidence, uses grilling to agree improvements to delivery speed, process, and codebase health, then applies selected documentation edits and follow-up ticket actions. Use when ending a substantial workstream or running a project retrospective.

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Prefer tightening or coupling with existing text.

    line 47
  2. Distinctive:

    Use both edit and ticket only when the guidance and its implementation are separately necessary.

    line 54
  3. Distinctive:

    Remove unsupported, session-specific, duplicate, and already-covered candidates.

    line 63
  4. Very specific:

    Run grilling one decision at a time with the boundary, evidence, current docs, absences, and candidates.

    line 65

≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Run retrospective

Assess the completed workstream through delivery-speed, process, and codebase-health lenses. The actual grilling skill owns the decision loop. Apply only documentation edits and ticket actions the user selects from the detailed summary.

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.
  • 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.

Route each lesson

  • Documentation edit: durable guidance in existing contracts, READMEs, docs/, ADRs, AGENTS, skills, templates, policies, or contributor guidance. Prefer tightening or coupling with existing text. Direct edits are limited to documentation.
  • Follow-up ticket: source, executable configuration, or other implementation is needed. Offer more detail, further grilling, normal or automatic create-issue, or skip; the delegated workflow retains its own gates.
  • Report only: useful evidence warrants neither an edit nor a ticket.

Use both edit and ticket only when the guidance and its implementation are separately necessary. A missing document may be created only when the user selects its exact proposed contents.

Workflow

  1. Resolve the workstream boundary and read relevant project documentation.
  2. Build an evidence ledger of outcomes, friction, rework, surprises, effective or missed gates, and successful practices under all three lenses.
  3. Remove unsupported, session-specific, duplicate, and already-covered candidates; classify the rest using the routes above.
  4. Run grilling one decision at a time with the boundary, evidence, current docs, absences, and candidates.
  5. Present the detailed summary:
    • findings under every lens, including no-finding results;
    • exact proposed documentation additions, replacements, or removals by file;
    • concise ticket summaries with all available action paths;
    • report-only observations, supporting evidence, confidence, and gaps.
  6. Obtain exact user selections through grilling. More detail or further grilling returns to that loop and regenerates the summary.
  7. Apply only selected documentation changes, preserving structure and avoiding duplication. Run only the selected ticket actions through create-issue.
  8. Compare the result with the confirmed action set, reread edited sections, and run declared documentation checks.
  9. Report changed docs, created issue links, report-only findings, confidence limits, and observed validation. Keep workstream history in chat.

Confirmation does not authorize commits, pushes, PRs, unselected issues, or other file edits. The original request to run a retrospective is not this final confirmation.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

Reviews can't be read right now.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.