Claude Code Plugins

Community-maintained marketplace

Feedback

self-learning-skills

@scottfalconer/self-learning-skills
20
0

Memory sidecar for agent work: recall before tasks, record learnings after tasks, review recommendations, optional backport bundles.

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name self-learning-skills
description Memory sidecar for agent work: recall before tasks, record learnings after tasks, review recommendations, optional backport bundles.

Self-learning sidecar

Use this skill to recall prior shortcuts before you start work, and to record durable “aha” moments + recommendations after you finish.

Critical rule: if no learnings exist (cold start), say so and proceed with standard tools — do not invent memories.

1) PRE-RUN: Recall (before starting work)

When to use: Before any non-trivial task.

Action:

  1. Locate the project store: <repo-root>/.agent-skills/self-learning/v1/users/<user>/
  2. Read <project_store>/INDEX.md (quick skim).
  3. If you need targeted recall, run:
    • python scripts/self_learning.py list --query "<keywords>"
    • Optional filters: --skill <name>, --tag skill:<name>
  4. Summarize 3–7 directly actionable bullets relevant to the current task (titles + IDs only; no long dumps).

2) POST-RUN: Record (after finishing work)

When to use: You discovered something durable (schema, fix, command sequence, constraint, etc.).

Action:

  1. Capture 1–5 Aha Cards (durable, reusable, specific, non-sensitive). Format: references/FORMAT.md.
    • Ensure every Aha Card and Recommendation has primary_skill (use unknown if unsure).
    • Set scope to project (repo/run-specific) or portable (generally reusable; a backport candidate).
    • If you rediscovered the same learning, treat it as reinforcement (signal) rather than duplicating the full card.
  2. Capture 1–5 concrete recommendations (what to change and where).
  3. Persist:
    • python scripts/self_learning.py record --json payload.json (or stdin)

Output requirement: print a short summary + top 3 items, then point to “view more” (INDEX.md / review --format json). Do not dump long JSON by default.

3) REVIEW: Dashboard / Next actions

When to use: “What’s still open?”, “What’s stale?”, “What should we backport?”, “Most useful learnings this week?”

Action:

  • python scripts/self_learning.py review --days 7
  • Full JSON: add --format json
  • Filters: --skill <name>, --scope project|portable, --status proposed,accepted,in_progress, --query "<keywords>"

4) MAINTENANCE / Governance

  • Repair store hygiene (append-only): python scripts/self_learning.py repair --apply
  • Update recommendation status/scope: python scripts/self_learning.py rec-status --id rec_... --status done --scope portable --note "..."
  • Optional backport bundle (explicit + auditable): python scripts/self_learning.py export-backport --skill-path <skill-dir> --ids <aha_ids> [--make-diff] [--apply]
  • Inspect backport markers in a skill: python scripts/self_learning.py backport-inspect --skill-path <skill-dir>

Docs

  • Setup/background: README.md
  • Integration templates (no hooks): references/INTEGRATION.md
  • Rubric/format/portability: references/RUBRIC.md, references/FORMAT.md, references/PORTABILITY.md