Plugin Dev Workflow Skill for Claude

Plugin Development Workflow

This repo is the Elixir/Phoenix Claude Code plugin. When editing plugin

files, follow this workflow to ensure quality.

Before You Start

Run make help to see all available commands:

make eval          # Quick: lint + score changed skills/agents
make eval-all      # Full: all 51 skills + 26 agents
make eval-fix      # Auto-fix + show failures
make test          # 220 pytest tests for eval framework and port tooling
make ci            # Full CI pipeline

Scoring Individual Files (CLI)

IMPORTANT: Always use -m module syntax, never run scorer.py directly.

# Score ONE skill (use -m, NOT direct file path)
python3 -m lab.eval.scorer plugins/elixir-phoenix/skills/verify/SKILL.md

# Score ONE skill with pretty output
python3 -m lab.eval.scorer plugins/elixir-phoenix/skills/verify/SKILL.md --pretty

# Score all skills
python3 -m lab.eval.scorer --all

# Score ONE agent
python3 -m lab.eval.agent_scorer plugins/elixir-phoenix/agents/verification-runner.md

# Score all agents
python3 -m lab.eval.agent_scorer --all
make ci            # Full CI pipeline

When Editing Skills (plugins/elixir-phoenix/skills/*/SKILL.md)

• Read CLAUDE.md conventions (size limits, frontmatter requirements)
• Make your changes
• Run make eval — it auto-detects changed skills and scores them
• If FAIL: check the dimension that failed, fix it
• Run make lint to verify markdown formatting
• Commit

Skill requirements (eval checks all of these):

• Frontmatter: name, description, effort. Description must start with action verb + include "Use when..."
• Iron Laws section with 1+ numbered items
• Under 185 lines (command skills) or 150 lines (reference skills)
• No section exceeds 45 lines
• All /phx: references point to existing skills
• All references/*.md paths exist
• No dangerous code patterns outside Iron Laws sections
• Code examples present (1+ fenced code blocks)
• "Use when..." in description (for trigger accuracy)

When Editing Agents (plugins/elixir-phoenix/agents/*.md)

• Make your changes
• Run make eval-agents to score all agents
• Agent requirements:
• permissionMode: bypassPermissions (always — background agents need it)
• disallowedTools: Write, Edit, NotebookEdit for review/analysis agents
• model matches effort: haiku=low, sonnet=medium, opus=high
• Under 300 lines (specialist) or 535 lines (orchestrator)

When Editing Eval Framework (lab/eval/*.py)

• Make your changes
• Run make test — all pytest tests must pass
• Run make eval-all — verify no skills/agents regressed
• If adding new matchers: add tests in lab/eval/tests/test_matchers.py

When Editing Hooks (plugins/elixir-phoenix/hooks/scripts/*.sh)

• Make your changes
• Run make lint (markdown in hook comments)
• Test the hook manually (hooks run on Edit/Write/Bash events)
• Check CLAUDE.md hook documentation is still accurate

Autoresearch (Self-Improvement Loop)

If make eval-fix shows failures, it suggests an autoresearch command:

# Copy-paste the suggested command from eval-fix output
claude -p 'Run autoresearch. Score all skills...' --allowedTools 'Edit,Read,Write,Bash,Glob,Grep'

This runs the autoresearch loop: find weakest skill → fix ONE issue → re-score → keep/revert.

Pre-Commit Checklist

Before committing any plugin changes:

• [ ] make lint passes
• [ ] make eval passes (changed files)
• [ ] make test passes (if eval framework changed)
• [ ] CHANGELOG.md updated (if user-visible change)
• [ ] Version bumped in plugin.json (if releasing)

References

• CLAUDE.md — full conventions, size limits, checklist
• lab/eval/ — scoring framework (24 matchers, 8 dimensions)
• lab/autoresearch/ — self-improvement loop
• lab/findings/interesting.jsonl — log interesting discoveries here