A Claude Code plugin providing skills for agentic engineering patterns and practices.
/plugin marketplace add 8thlight/lightfactory
/plugin install lightfactory@praxis
After installation, skills are available as praxis:skill-name and activate automatically when relevant to your task.
git clone https://github.com/8thlight/lightfactory/plugin marketplace add /path/to/lightfactory
/plugin install lightfactory@praxis
/plugin marketplace update lightfactory@praxis
| Skill | Command | Description |
|---|---|---|
| research | /research |
Spawns parallel subagents to explore a codebase and produce a compact research artifact |
| plan-tasks | /plan-tasks |
Consumes research artifact and produces a compact implementation plan with L3/L4 test specs |
| implement | /implement |
Executes an implementation plan phase by phase with strict test-first discipline |
| tdd | /tdd |
Boundary-focused TDD workflow enforcing L3/L4 altitude testing and property-based tests |
| harness | /harness |
Audits the agentic coding environment across nine pillars and produces a gap analysis with actionable improvements |
| adr | /adr |
Guides writing minimal Architecture Decision Records |
| reflect | /reflect |
Post-session reflection that mines git history and artifacts to produce improvement proposals |
The praxis plugin's core workflow for non-trivial features follows three phases:
- Research (
/research) — Explore the codebase with parallel subagents, output a compact research artifact to.light/sessions/ - Plan (
/plan-tasks) — Consume the research artifact, produce a compact implementation plan with test specs and Agent Context blocks - Implement (
/implement) — Execute the plan phase by phase with strict RED → GREEN → VALIDATE discipline
- Plan Mode Native — The workflow aligns with Claude Code's native plan mode. Research happens outside plan mode; planning activates draft behavior inside plan mode; implementation executes the approved plan.
- Three-Agent TDD Isolation — During implementation, each TDD phase dispatches isolated agents:
agent-test(writes failing tests),agent-impl(writes minimal code to pass), andagent-validate(runs full test suite). Agents never modify each other's artifacts. - Context Compaction — Research produces a ~200-line artifact that carries forward into planning, replacing unbounded codebase exploration with a focused summary. Plan mode entry is the compaction boundary.
- Tracker Detection Chain — The implement and plan-tasks skills auto-detect available task trackers: yaks (preferred) → beads (fallback) → native tasks (last resort).
- TDD (
/tdd) — For interactive, human-in-the-loop test-driven development outside the full RPI flow. Enforces boundary-focused testing at the L3/L4 altitude with ZOMBIES progression. - ADR (
/adr) — Guides writing Architecture Decision Records following the Harmel-Law signal check pattern. Ensures decisions are actually made before documenting them. - Harness (
/harness) — Audits the agentic coding environment (instruction files, hooks, type safety, linting, pre-commit, architecture tests, CI gates, sandbox, secret scanning) and implements selected improvements. Based on Fowler's feedforward/feedback control model. - Reflect (
/reflect) — Post-session learning loop that mines git history and session artifacts to produce improvement proposals for skills, CLAUDE.md, and hooks.
Skills are validated by a three-layer pipeline. See tests/README.md for full methodology.
Layer 1 — Deterministic (local): Validates skill structure, frontmatter, triggers, and scenario schemas. No API calls.
bash tests/local/validate-skills.shLayer 2 — Promptfoo evals: Functional and behavioral evaluation using real Claude API calls. See tests/evals/README.md for setup and cost.
cd tests/evals && promptfoo evalLayer 3 — Human review: Manual review for subjective quality and calibrating LLM-judge rubrics.
See AGENTS.md for skill authoring guidelines, testing workflows, and the pre-ship checklist.