agentleFS
subarnasaikiasubarnasaikia/kaggle-skills/.claude/skills/retrospect-session/SKILL.md

retrospect-session

Skill2 starsChanged 6 months agoMITView on GitHub

Session-end retrospective that extracts 0-3 durable learnings from the conversation, updates the LEARNINGS.md digest, and flags recurring scars. Use when the user says "retro", "retrospect", "wrap up", "end of session", "what did we learn today", or when a long working session is winding down. The Stop hook also nudges this skill.

We split this skill into the rules it gives an agent

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

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    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:

    Keep pinned ≤ 15 entries. Evict the oldest non-recurring high-cost ones when over budget.

    line 24
  2. Very specific:

    Prune stale: If any learning references code paths that no longer exist, a Kaggle API behavior that has changed, or a library version no …

    line 27

≈ “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

Session Retrospective

Reviews what happened during the session, extracts durable learnings, and keeps LEARNINGS.md lean.

Steps

  1. Scan the session. Identify moments where:
    • An approach was tried and failed.
    • An approach was tried and worked against expectation.
    • A tool / API / environment quirk was discovered.
    • A user correction changed direction.
    • A hypothesis was confirmed or rejected by data.
  2. Filter to durable material. Drop anything already recorded in code, git history, a commit message, or an existing learning. Target: 0 to 3 new learnings per session; zero is a valid answer.
  3. For each surviving item, invoke the capture-learning skill (one per item).
  4. Update LEARNINGS.md digest:
    • Pull the 10 most recent scar + pattern learnings across all competitions.
    • Render them as - L-… — <Rule> (scope, cost) under the <!-- BEGIN:RECENT --> block.
    • For any learning with recurring: true OR cost: high, also ensure it appears in the <!-- BEGIN:PINNED --> block (dedupe by ID).
    • Keep pinned ≤ 15 entries. Evict the oldest non-recurring high-cost ones when over budget.
    • Alternatively, run scripts/kln digest to rebuild both blocks automatically.
  5. Detect recurrence. For any newly-captured scar, run grep -l -i "<same root cause>" .learnings/L-*.md. If the same root cause has appeared ≥ 2 times, mark recurring: true in the new file AND the older file(s), and add to pinned.
  6. Prune stale. If any learning references code paths that no longer exist, a Kaggle API behavior that has changed, or a library version no longer in use, move it to .learnings/archive/ with a one-line reason in its frontmatter (archived_reason: ..., archived_date: ...). Never silently delete.
  7. Write a 3-line summary to the user:
    Retro:
      Captured: L-…, L-…
      Pinned: L-… now recurring
      Archived: L-… (API deprecated)

When there's nothing to capture

Say so in one line: "Retro: no durable learnings this session." Do not invent. Do not pad. Zero learnings is compatible with a productive session.

Creative directive — the scar compiler

Think of yourself as editing a project post-mortem wiki. Every line should be readable by a stranger. Rules first, war stories second. If a learning file reads like a diary entry, rewrite the Rule line so the future reader doesn't have to unpack the story.

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.