coach-plugin

Coach Creation Plugin

A Claude Code plugin that scaffolds domain-specific AI coaches from a single /create-coach command.

You answer 4 questions. It generates a complete coaching environment: a local knowledge base built from your reference materials, a persistent subject profile, a session log, and a structured coaching skill that reads the log, searches the library, and produces cited recommendations.

Everything runs locally. No data leaves your machine.

What You Get

Each generated coach project contains:

ai-{domain}/
├── CLAUDE.md                  # Subject profile + coach behavior rules
├── log.md                     # Append-only session log
├── .claude/skills/{domain}.md # /{domain} coaching skill
├── .mcp.json                  # Connects qmd search to Claude Code
├── scripts/
│   ├── convert.py             # Resources → Markdown (docling)
│   ├── index.sh               # Build/update search index (qmd)
│   └── qmd-mcp.sh             # MCP server wrapper
├── resources/                 # Drop your reference materials here (any structure)
├── context/                   # Auto-generated Markdown (flat)
├── .cache/                    # Local search index (qmd + embeddings)
└── .venv/                     # Python env (docling)

How It Works

The coaching loop follows a rigid 6-step structure regardless of domain:

1. Read the log — load all entries from the current phase
2. Identify the issue — what the subject is reporting
3. Search the library — query the local knowledge base for evidence
4. Synthesize — combine log data with library findings
5. Cite the source — name the book/paper backing the recommendation
6. Log the change — append the recommendation to the log

The structure is fixed. The domain-specific content (metrics, failure modes, search terms, phases) is generated from your interview answers during setup.

Requirements

• qmd on PATH — on-device search engine

  npm install -g @tobilu/qmd
  

• uv on PATH — Python environment manager

  curl -LsSf https://astral.sh/uv/install.sh | sh
  

Install

git clone https://github.com/EdwardAstill/coach-plugin.git
claude plugin marketplace add ~/path/to/coach-plugin
claude plugin install coach-plugin

To install for the current project only, use --scope local:

claude plugin install coach-plugin --scope local

Usage

From any Claude Code session:

/create-coach

The wizard asks 4 questions:

1. Domain — what is this coach for? (nutrition, learning, rehab, music practice, etc.)
2. Goals — what are your goals and current level?
3. Constraints — anything the coach must never ignore?
4. Anything else — philosophy, metrics, preferences?

Then it suggests a project location, scaffolds everything, sets up the Python environment, and optionally runs the ingestion pipeline if you've already dropped resources in.

Adding Knowledge

Drop reference materials into resources/ (any subfolder structure you like), then run:

.venv/bin/python3 scripts/convert.py && bash scripts/index.sh

Or just ask Claude to do it for you.

Domain Examples

This plugin has been used to create coaches for:

• Strength training — periodization, RPE autoregulation, fatigue management
• Nutrition — macro targets, meal adherence, refeeding protocols
• Learning — spaced repetition, deliberate practice, retention tracking
• Rehab — graded exposure, pain science, load tolerance

Any domain where you have reference material and want structured, evidence-based coaching works.