Portable customizations for AI coding agents.
Today that's a set of Agent Skills under skills/, plus the subagents one of them dispatches; prompts and themes may follow.
Each skill is a self-contained directory holding a SKILL.md (frontmatter + instructions) and any supporting files.
The format is agent-agnostic: Pi, Claude Code, and OpenAI Codex all discover SKILL.md directories.
A skill's maturity is its directory.
skills/stable/ is settled: prompts and names change only with a version bump.
Everything under skills/experimental/ is alpha: prompts, formats, and names may change without notice.
A skill graduates from one to the other once it has proven itself through real use.
| Skill | What it does |
|---|---|
stable/pr | Open a pull/merge request on the repo's forge (GitHub, GitLab, Gitea, Forgejo) using the forge's native CLI. |
experimental/intent | Capture an idea as a version-controlled intent — the first artifact in the intent -> spec -> plan chain. |
experimental/spec | Turn an accepted intent into a spec: what the solution must do, checked against organizational policy. |
experimental/plan | Turn an approved spec into a read-only implementation plan — the audit trail and review baseline. |
experimental/execute | Run an approved plan: a worker subagent implements, an adversarial judge reviews, and the loop ends only when the judge finds nothing. |
intent, spec, and plan form the SDLC artifact chain — one numbered chain of reviewed markdown per change, with the stage gates enforced by a hook rather than by asking the model.
execute is what happens after the chain: it ships two subagents, worker and judge, and the judge reviews in a context that never sees the worker's account of its own work.
See docs/sdlc.md for a worked example.
npx skills add lstig/agentsAdd -g to install globally instead of per-project, --skill <name> to pick individual skills, or --list to see what's available.
Works across Claude Code, Cursor, Codex, Copilot, Gemini CLI, and Cline.
The repo is a Claude Code plugin marketplace (.claude-plugin/marketplace.json) named lstig-agents, offering two plugins:
development—pr, for opening a pull or merge request on any forge.sdlc— the artifact chain:intent,spec,plan,executeand its two subagents, plus thePreToolUsehook that enforces the gates.
/plugin marketplace add lstig/agents
/plugin install development@lstig-agents
/plugin install sdlc@lstig-agents
Point your agent's skills directory at these, or copy individual skills into your own dotfiles.
This route carries skills only: execute arrives without worker and judge, and stops rather than doing the work itself.
For example, to install one for Claude Code:
git clone https://github.com/lstig/agents.git
cp -R agents/skills/stable/pr ~/.claude/skills/prPi and Codex read the same SKILL.md layout; see your agent's docs for its skills location.
Vendoring intent, spec, or plan gets the skills but not the gate hook; bring .claude-plugin/hooks/ too, or the gates become instructions rather than enforcement.