Curated skills for AI coding agents (Claude Code, Codex, OpenCode).
lefant/- Custom skills developed in-housevendor/- Vendored skills from upstream repositories (reviewed for security)
- Adopt Documentation & Process — A single-file instruction for adding a documentation map and development workflow to another repository's README, with pointers from AGENTS.md and CLAUDE.md.
Use these documentation areas. Create optional directories only when there is content to own; not every area below exists in this tooling repository yet.
docs/specs/: Feature Specifications — living documents describing WHAT features do and WHY- Skill:
feature-specs
- Skill:
docs/decisions/: Architecture Decision Records (ADRs) — immutable records of architectural choices once accepted; supersede rather than rewrite their rationaledocs/changelog/: Release notes and changelog fragments- Skill:
changelog-fragments
- Skill:
docs/solutions/: Documented solutions to past problems and team learnings, organized by category with YAML frontmatter (module,tags,problem_type)docs/reference/: Technical references and operational detailsdocs/api/: Technical references for internal APIs, when applicabledocs/brand-guide.md: Brand and design guide, when applicabledocs/research/: Discovery notes and analysisdocs/brainstorms/: Existing standalone brainstorms and design explorationsdocs/plans/: Time-stamped implementation plansdocs/devlog/: Time-stamped implementation outcomes, caveats, and learnings- Skill:
devlog
- Skill:
The spec, ADR, and changelog skills replace the old spec_create, adr_create,
and changelog_create commands. Use the installed agent's skill invocation syntax;
the links above point to the maintained skill instructions.
Workflow: Research/Plan → Implement → Review → Compound
Prefer the Compound Engineering skills when installed. Verify the installed version's skill names and output paths before invoking them; this repository does not install that plugin. The workflow also works without it.
Commit progress continuously in small, reviewable increments. Once continuous pushes to the agreed remote and branch are authorized, push each coherent, checked checkpoint without asking again. Inspect automatic deployment effects first; separate release and shared-state actions still need their own authorization.
- Research/Plan — document current behavior, constraints, and requirements;
define implementation strategy and verification. Durable research belongs in
docs/research/, and plans indocs/plans/. A plugin may keep brainstorm and planning work in one plan artifact; do not duplicate it just to fill folders. - Implement — make the source changes and verify them against the plan.
- Review — inspect correctness, contracts, security, and test coverage; resolve findings and rerun affected checks.
- Compound — capture verified reusable solutions in
docs/solutions/and update their existing owners rather than duplicating knowledge.
At meaningful checkpoints, use devlog to record outcomes,
evidence, caveats, and follow-ups in docs/devlog/. Devlogs preserve session history;
solutions preserve reusable knowledge. Use atomically-land
when closing out work that also needs plans, specs, ADRs, or changelog updates.
Use docs/research/, docs/plans/, and docs/devlog/ for new work. The
rpi skill uses these documentation destinations while
preserving its explicitly requested legacy workflows; it is not the default
process. Historical documents may retain former thoughts/shared/ paths as
provenance. Do not move historical files merely to follow the new convention.
skills add /path/to/agent-skills --skill '*' -g -y -a claude-code -a codex -a opencodeWildcard installs include optional packages. Caveman, ZFC, and Pulumi are
explicit opt-in / repository-optional,
not global defaults; select individual skills instead of '*' to omit them.
Cloned to /opt/lefant-agent-skills and symlinked via entrypoint.
./scripts/update-vendor.shThen review changes before committing.
For a targeted refresh, pass upstream repository names, for example:
./scripts/update-vendor.sh antithesishq/antithesis-skills typesafe-ai/skillsThe updater stages fetches, local patches, and checks before replacing vendor/.
Failures leave the existing snapshot unchanged. See the
vendor maintenance notes for held updates.
Before committing vendor changes, verify skill discoverability:
./scripts/check-vendor-layout.sh
python3 scripts/tests/vendor-safety.test.py -vFor quick manual inspection:
find vendor -type d -name skills -print
find vendor -path '*/SKILL.md' -printf '%h\n' | sortRules:
find vendor -type d -name skills -printshould normally print nothing. Vendored skills must be flattened tovendor/<source>/<skill>/SKILL.md, even when upstream stores them underskills/.- Do not vendor whole plugin bundles that need agents, commands, MCP config, or other non-skill assets unless the installer can consume those assets. Handle plugin-native bundles in a plugin/package layer instead.
- Keep source-to-local path mappings in
scripts/update-vendor.sh; do not rely on symlinks for discoverability. - Keep frontmatter
name:values unique acrosslefant/andvendor/; installers expose one skill per name.
| Skill | Description |
|---|---|
github-access |
Access GitHub repositories via gh CLI or REST API |
sentry |
Fetch and analyze Sentry issues, events, and logs |
architecture-decision-records |
Create and manage ADRs |
changelog-fragments |
Maintain changelog fragments for conflict-free history |
feature-specs |
Document feature requirements and scenarios |
atomically-land |
Close out implementation work and update surrounding docs/tracking |
devlog |
Write implementation devlogs for current session state |
exa |
Use Exa for web search, code context lookup, and current-information lookups via SDK, API, or MCP |
skills-best-practices |
Create, review, and refactor agent skills with self-contained packaging and eval guidance |
github-get-pr-comments |
Gather PR comments and combine them with local project context |
git-resolve-merge-conflicts |
Resolve git merge conflicts safely using local context |
mermaid-diagrams |
Create hierarchical Mermaid diagrams |
handover |
Generate a concise resume prompt for the next session |
hitl-browser-automation |
Run human-in-the-loop browser automation with VNC, loopback CDP, smoke validation, and agent-browser continuation |
rpi |
Preserve the create_plan, implement_plan, and research_codebase workflows |
qrspi |
Portable adaptation of matanshavit/qrspi's eight-phase workflow; upstream provenance and local changes documented in lefant/qrspi/SOURCE.md |
recent-context-from-git |
Summarize recent local docs and work context from git history |
test-analyzer |
Analyze CTRF test reports with jq |
expect-testing |
Create and review expect tests, inline snapshots, approval tests, and golden files |
youtube-transcript |
Fetch YouTube video transcripts |
tasknotes |
Manage Obsidian tasks through maintained TaskNotes interfaces or a safe Markdown fallback |
untis-access |
Access WebUntis data, compare repo implementations, and run proven message/timetable scripts |
See vendor/README.md for sources and update process.