🧪 A Claude Code plugin that analyzes any codebase and generates a tailored skill pack — coding skills, review agents, Q&A agents, and a shared guidelines document — then keeps them evolving through continuous learning.
Agent
Devil's advocate agent that argues against each finding before it is accepted into project learnings. Verifies claims against the codebase, searches for counterexamples, assesses blast radius, checks for contradictions with existing guidelines, and renders accept/reject/modify verdicts. This agent is invoked by the…
Agent
Scans a codebase for all existing documentation, coding standards, and developer instructions. Discovers AI agent rules (Cursor, CLAUDE.md, Copilot), architecture decision records, style guides, convention configs, and workflow documentation. Returns a structured documentation profile. This agent is invoked by the…
Agent
Compares current project guidelines against actual codebase patterns to detect drift. Extracts discrete claims from guidelines, greps the codebase for evidence and counter-evidence, calculates adherence ratios, and flags divergences. Returns a structured drift report as JSON. This agent is invoked by the potion-learn…
Agent
Analyzes git history, branch structure, merge patterns, and PR conventions to extract the team's workflow practices. Uses git CLI and optionally GitHub/GitLab CLI for PR template and review process analysis. Returns a structured git workflow profile as JSON. This agent is invoked by the potion-skill-generator during…
Agent
Synthesizes challenged findings into the .claude/learnings.md staging file and prepares merge proposals for guidelines. Handles ID assignment, deduplication against existing learnings and archives, and archive management when merges are approved. This agent is invoked by the potion-learn skill during Phase 3 (Write) …
Agent
Performs a deep-dive into a single codebase module to extract patterns, conventions, and architecture. Returns a structured module profile as JSON. This agent is invoked by the potion-skill-generator during Phase 2 — not meant for direct use. Covers auth modules, data pipelines, component libraries, API services…
Agent
Analyzes a codebase's structure to identify distinct modules, services, and packages. Returns a structured module map as JSON. This agent is invoked by the potion-skill-generator during Phase 1 — not meant for direct user use. Handles monorepos (Turborepo, Nx, Lerna), Django modular-monoliths, Rust workspaces with…
Agent
Synthesizes module profiles into a unified guidelines.md document. Identifies cross-cutting patterns, resolves inconsistencies, and produces the shared knowledge base for all generated skills. This agent is invoked during Phase 3. Handles consistent monorepos, mixed-language microservice architectures, and…
Agent
Fetches and classifies review comments from a single pull request to extract actionable conventions. Returns structured findings as JSON. This agent is invoked by the potion-learn skill during Phase 1 (Gather) — not meant for direct use. Filters bot comments, classifies human feedback into convention categories, and…
Agent
Mines merged pull request review comments from GitHub or GitLab to extract tribal knowledge: enforced conventions, common mistakes, architectural preferences, and testing expectations. Returns a structured review-patterns profile as JSON. This agent is invoked by the potion-skill-generator during Phase 2, running in…
Agent
Extracts cross-cutting conventions shared across all language stacks in a monorepo: git workflow, CI/CD, deployment, shared types/contracts, and observability. Produces shared.md — the foundation that per-stack synthesizers reference. Invoked during Phase 3 in multi-stack mode only.
Agent
Generates production-ready Claude Code skills and subagents from guidelines and module profiles. Produces the complete skill pack. This agent is invoked during Phase 4. Creates ask, plan, implement, and review skills plus explorer, planner, implementer, and reviewer agents — all tailored to the target project's…
Agent
Synthesizes module profiles for a single language stack into a complete guidelines directory. Produces stack-specific patterns, conventions, testing, and pitfalls. References shared.md for cross-cutting conventions to avoid duplication. Invoked during Phase 3 in multi-stack mode only — one instance per stack, all…
Agent
Normalizes free-form text (CodeRabbit exports, meeting notes, team decisions, linter discussions) into structured convention findings. Returns JSON in the same format as the PR Miner. This agent is invoked by the potion-learn skill during Phase 1 (Gather) — not meant for direct use.