Skip to content

About

No description, website, or topics provided.

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

138 Commits

Folders and files

Repository files navigation

lefant/agent-skills

Curated skills for AI coding agents (Claude Code, Codex, OpenCode).

Structure

  • lefant/ - Custom skills developed in-house
  • vendor/ - Vendored skills from upstream repositories (reviewed for security)

Reference Guides

  • 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.

Documentation & Process

Documentation (./docs)

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
  • docs/decisions/: Architecture Decision Records (ADRs) — immutable records of architectural choices once accepted; supersede rather than rewrite their rationale
  • docs/changelog/: Release notes and changelog fragments
  • 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 details
  • docs/api/: Technical references for internal APIs, when applicable
  • docs/brand-guide.md: Brand and design guide, when applicable
  • docs/research/: Discovery notes and analysis
  • docs/brainstorms/: Existing standalone brainstorms and design explorations
  • docs/plans/: Time-stamped implementation plans
  • docs/devlog/: Time-stamped implementation outcomes, caveats, and learnings

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.

AI-Assisted Development (./docs)

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.

  1. Research/Plan — document current behavior, constraints, and requirements; define implementation strategy and verification. Durable research belongs in docs/research/, and plans in docs/plans/. A plugin may keep brainstorm and planning work in one plan artifact; do not duplicate it just to fill folders.
  2. Implement — make the source changes and verify them against the plan.
  3. Review — inspect correctness, contracts, security, and test coverage; resolve findings and rerun affected checks.
  4. 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.

Usage

With skills CLI

skills add /path/to/agent-skills --skill '*' -g -y -a claude-code -a codex -a opencode

Wildcard 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.

In Docker

Cloned to /opt/lefant-agent-skills and symlinked via entrypoint.

Updating Vendored Skills

./scripts/update-vendor.sh

Then review changes before committing.

For a targeted refresh, pass upstream repository names, for example:

./scripts/update-vendor.sh antithesishq/antithesis-skills typesafe-ai/skills

The updater stages fetches, local patches, and checks before replacing vendor/. Failures leave the existing snapshot unchanged. See the vendor maintenance notes for held updates.

Vendor Quality Checks

Before committing vendor changes, verify skill discoverability:

./scripts/check-vendor-layout.sh
python3 scripts/tests/vendor-safety.test.py -v

For quick manual inspection:

find vendor -type d -name skills -print
find vendor -path '*/SKILL.md' -printf '%h\n' | sort

Rules:

  • find vendor -type d -name skills -print should normally print nothing. Vendored skills must be flattened to vendor/<source>/<skill>/SKILL.md, even when upstream stores them under skills/.
  • 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 across lefant/ and vendor/; installers expose one skill per name.

Skills Included

Custom Skills (lefant/)

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

Vendored Skills (vendor/)

See vendor/README.md for sources and update process.

About

No description, website, or topics provided.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages