Self-Learning Skills
Teach your AI coding agent to recognize hard-won golden paths in a session and persist them as reusable skills for next time.
Least privilege: writes only skill/rule files and notes; README scopes this clearly, 2. User confirmation: README states capture happens 'no prompt needed' with post-hoc notification — writes are default-unconfirmed, 1. Data flow: destination table is clear (skills/, .cursor/rules/learned/, AGENTS.md), 2. Sensitive data: explicit never-write-secret-values design recording only locations is a strength, but it is a stated policy with no shown mechanism, 2. Dependency security: install relies on community skills CLI via npx remote execution with no pinning/audit note, 1. External effects: modifies user config files and content gets committed/shared; acknowledged but unguarded, 2. Rollback: no undo/removal instructions for harvested skills at all, 0. Source attribution: no labeling of where harvested knowledge originated; publisher unverified, 1.
Self-consistency: the loop, triage, and promotion rule cohere, 2; however the core SKILL.md was not present in the provided evidence, so cross-checking was impossible — the deduction is reflected in low confidence. Dependency availability: depends on vercel-labs skills CLI and each agent's rule-loading with no version/availability commitments, 1. Failure messages: nothing on misfired triggers, write failures, or triage misrouting feedback, 1.
Audience and scenarios: explicitly targets developers on Claude Code/Cursor/AGENTS.md agents with concrete debugging and ops-recurrence scenarios, 3. Capability boundaries: triage and promotion thresholds are specified, but when the meta-skill itself fails is not, 2. Trigger precision: concrete recognition cues are given, but detection is left to LLM judgment with no hard guardrails, 2. Environment fit: per-tool persistence paths plus manual install, 3.
Information architecture: layout diagram is clear and well-layered, but the actual SKILL.md etc. were not in evidence — README description only, 2. Install notes: npx/plugin/manual all present with commands, 3. Naming stability: self-learning naming consistent; learned/ convention explicit, 2. Examples and FAQ: only prose on the promotion rule; no worked capture example or FAQ, 1. Known limitations: no dedicated limitations section, 1. License: full MIT text present, 3. Versioning/changelog: no version number or CHANGELOG, 1. Maintenance responsibility: copyright holder only; no maintenance statement or update path, 1.
Output usability: produces standard SKILL.md/.mdc/AGENTS.md entries with a template and authoring reference, directly loadable, 2. Marginal value: addresses a real pain point; failure-pattern recording and the promotion rule are distinctive value, 2. Cost benefit: lightweight, no runtime deps, triage limits bloat; misdirected auto-writes require manual cleanup, 2.
Claim traceability: install paths and layout are checkable, but core behavioral claims (recognition, never writing secrets) have no code evidence in the review set, 1. Cross-source corroboration: only README and LICENSE were provided; SKILL.md, AGENTS.md, skills.sh. absent, so no mutual verification, 1. Fact/inference separation: the promotion rule is flagged as community-suggested, but marketing language ('70+ agents') is blended with factual claims, 1.
- The skill writes config files and commits shared content without user confirmation by default; review its write targets before production use and establish your own rollback process — the repo provides no undo mechanism.
- 'Never writes secret values' is a README claim only; the core SKILL.md was not part of the reviewed evidence. Do not assume a mechanical leak guard — manually inspect harvested content for tokens and connection strings before committing.
- The recommended install path pulls a third-party CLI via npx, a supply-chain risk; prefer manual copy and pin versions.
- Publisher is unverified and the repo has no version number, changelog, or maintenance commitment — assess long-term viability before adoption.
What does this agent do, and when should you use it?
self-learning-skills is a self-improving meta-skill for AI coding agents, working with Claude Code, Cursor, and any agent that reads an AGENTS.md file. It addresses a concrete problem: commands, credential locations, and deploy workflows figured out during hard debugging evaporate when a session ends, forcing the next session to start from zero. The project teaches the agent to recognize the moment it has earned a reusable golden path and persist it where the tool auto-loads it — a new skills/<name>/SKILL.md for Claude Code, or .cursor/rules/learned/<name>.mdc for Cursor. It captures not just the working path but the failures, since skipping a known dead-end next session is often worth more than the win itself. Triage routes each lesson to a skill, a lightweight memory note, or a skip, and a promotion rule requires a passing check, a named failure pattern, and at least one ruled-out dead-end before a session becomes an authoritative skill. Harvested rules are designed to never contain secret values, only pointers to where secrets live.
The meta-skill does no coding itself; instead, it drives the host agent through a harvest loop: 1) Recognize the moment — a task that only worked after several tries, a non-obvious command, a newly learned project fact, a recurring operational workflow, or the user saying 'remember this'. 2) Capture immediately — it picks the scope and name itself, records the procedure (not a one-off answer) plus a 'what didn't work' note. 3) Reuse — next session the entry loads automatically via skill description matching or an always-read instructions file. Persistence differs per tool: Claude Code and Codex write a new skills/<name>/SKILL.md; Cursor writes .cursor/rules/learned/<name>.mdc; Zed, Aider, Gemini CLI and similar append to AGENTS.md. The repository ships AGENTS.md (cross-tool generic version), skills/self-learning/SKILL.md (the recognize-and-harvest procedure), references/skill-authoring.md (a condensed authoring spec), assets/SKILL.template.md (a fill-in template), plus a Claude Code plugin manifest and a skills.sh registry manifest.
- A Claude Code developer repeatedly re-learning across sessions how to reach the prod DB, where creds live, and what the deploy command is.
- A Cursor user who wants harvested experience rules shared team-wide via the .cursor/rules/ directory in git.
- Users of AGENTS.md-reading agents like Codex, Zed, Aider, or Gemini CLI who want lessons persisted into an always-read instructions file.
- An engineer maintaining many projects who installs once globally (npx skills add -g) so every project shares the harvest mechanism.
- A cautious user worried about unverified guesses being enshrined as skills — the promotion rule requires a passing check, a named failure pattern, and a ruled-out dead-end.
What are this agent's strengths and limitations?
- Cross-tool by design: one loop adapts to Claude Code, Cursor, Codex, and any AGENTS.md-reading agent, keeping knowledge portable between tools.
- Captures what didn't work: named failure patterns and ruled-out dead-ends let the next session skip known dead-ends outright.
- The promotion rule prevents bloat and guess-enshrinement: a skill only forms with a passing check, a named failure pattern, and at least one ruled-out dead-end.
- Safety built in: harvested skills never contain passwords, tokens, connection strings, or API keys — only pointers like env var names, selector functions, or MCP tools.
- It is a meta-skill, not an executor: its value depends entirely on the host agent's skill/rule loading mechanism, requiring a supported runtime like Claude Code or Cursor.
- Harvest quality depends on the agent's judgment — the README offers no benchmarks or evaluations of recognition and triage accuracy.
- Harvested skills/rules get committed and shared; lessons harvested in one team's workflow may not fit another's projects.
- Cursor-harvested rules accumulate in .cursor/rules/learned/, requiring periodic review to avoid rule-file sprawl.
How do you install or deploy this agent?
Recommended via npx (community skills CLI, auto-detects 70+ agents):
- Install into detected agents: npx skills add kulaxyz/self-learning-skills
- Global: npx skills add kulaxyz/self-learning-skills -g
- Specific agent: npx skills add kulaxyz/self-learning-skills -a claude-codeClaude Code plugin:
- /plugin marketplace add kulaxyz/self-learning-skills
- /plugin install self-learning@self-learning-skillsManual: git clone https://github.com/kulaxyz/self-learning-skills, then for Claude Code run cp -R self-learning-skills/skills/self-learning ~/.claude/skills/ (global) or copy into a project's .claude/skills/; for Cursor run mkdir -p .cursor/rules && cp self-learning-skills/.cursor/rules/self-learning.mdc .cursor/rules/; for AGENTS.md agents run curl https://raw.githubusercontent.com/kulaxyz/self-learning-skills/main/AGENTS.md >> AGENTS.md. No credentials or API keys are required.
How do you use this agent?
After installation there is nothing to trigger manually: the agent recognizes harvest moments during a session (multi-try tasks, non-obvious commands, recurring operational workflows, or you saying 'remember this') and acts immediately, telling you afterward what it persisted. To try once without installing: npx skills use kulaxyz/self-learning-skills --skill self-learning | claude. Each lesson is triaged — multi-step reusable procedures become a skill/rule, single facts go to lightweight notes (e.g., a MEMORY.md), and genuine one-offs are skipped. Next session, skills load automatically by description matching, and AGENTS.md entries apply because the file is always read.