Faf Expert

FAF Expert — Master the Format

.faf is an IANA-registered context format (application/vnd.faf+yaml) — a typed, portable file you own, readable by any AI, with no bespoke manifest and no vendor lock-in. This is the mechanic's manual: scoring internals, MCP configuration, bi-directional sync, and the full slot model.

The FAF skill family — where this fits

New to FAF? Start with faf-context to reach 100%. Come here to go deep.

When to use this skill

• Configuring .faf files and the MCP server beyond the basics
• Understanding the scoring engine and how to drive a project to high tiers
• Multi-AI workflows — one context across Claude, Cursor, Gemini, Codex, Windsurf
• Reviving a legacy codebase into AI-readable project DNA
• Standardizing context format across a team

The format

project/
├── package.json     ← dependencies (npm reads this)
├── project.faf      ← AI context (any AI reads this)
├── CLAUDE.md        ← human docs (synced from .faf)
└── src/             ← code (guided by the context)

README.md is prose for humans · CLAUDE.md is prose for Claude · project.faf is structure for ANY AI (IANA application/vnd.faf+yaml). One source of truth; everything else is emitted from it.

Universal compatibility — one .faf, every tool

Quick start

1. Install the CLI

npm install -g faf-cli        # or: brew install wolfe-jam/faf/faf-cli
faf --version

2. Configure the MCP server — add to Claude Desktop's claude_desktop_config.json, then restart Claude:

{
  "mcpServers": {
    "faf": {
      "command": "npx",
      "args": ["-y", "claude-faf-mcp@latest"]
    }
  }
}

3. The loop to 100% ✪

faf auto     # 1. detect stack + language, seed context from your README
faf score    # 2. the % + exactly which slots are still empty
faf go       # 3. guided fill — answer only the gaps the AI can't source
faf score    # 4. 100% ✪
faf sync     # 5. push context → CLAUDE.md / AGENTS.md

CLI commands

# The path to Trophy
faf auto                # detect + seed (fills what it can)
faf score               # AI-readiness % + the empty slots
faf go                  # guided fill — close the gaps
faf sync                # emit CLAUDE.md / AGENTS.md from the .faf

# Anytime
faf init                # create a fresh project.faf
faf check               # validate structure (IANA-clean)
faf git <url>           # score any GitHub repo → .faf, no clone
faf formats             # browse the supported stacks
faf migrate             # upgrade an older .faf
faf export --copilot    # emit .github/copilot-instructions.md
faf taf setup           # wire up TAF test receipts (CI)

MCP server tools

claude-faf-mcp brings the format into the tool surface. The live tool set (verified against the running server):

# The path to 100% ✪
faf_auto     → scan manifests (package.json, Cargo.toml, pyproject.toml…) + fill the .faf
faf_score    → 0–100% AI-readiness + tier + per-slot breakdown
faf_go       → the friendly front door — asks the 6 Ws
faf_sync     → sync .faf → CLAUDE.md (+ optionally AGENTS.md)

# Anytime
faf_init     → create a new project.faf (name, goal, language) + starting score
faf_context  → set / show the active project path faf_ calls resolve against
faf_doctor   → diagnose: empty / weak slots + how to fix each
faf_enhance  → refine the .faf with an AI model (claude / gemini / grok, optional consensus)
faf_trust    → attest integrity: validity, score + a deterministic parity hash anyone can verify
faf_bench    → prove the .faf earns its place — the context's worth on THIS repo, falsifiably
faf_about    → explain the FAF format — project DNA for AI (IANA-registered)

# Project soul (.fafm) — memory across sessions
faf_etch     → remember a decision / gotcha / win to the .fafm
faf_recall   → recall memories from the .fafm, ranked by priority + recency

(The exact set evolves release to release — verify against your installed version with the MCP client's tool list.)

Scoring

Tiers

The score is deterministic — a WASM-compiled engine, mechanical and falsifiable. Same input → same score, every time. FAF doesn't lie.

The 21-slot model

faf-cli scores on 21 slots. Your app_type selects which are active — a CLI ignores frontend slots (slotignored, never counted against you). Detection fills the stack + language; you supply only the underivable bits — project.name/goal + the 6 Ws (a sharp goal seeds several). 100% = every active slot filled (not all 21). (Teams / Enterprise tiers add more slots — separate from faf-cli's 21.)

# project.faf — IANA application/vnd.faf+yaml
faf_version: "3.0"

project:                          # required
  name: modern-web-app
  goal: Real-time collaboration platform for distributed teams
  main_language: typescript

human_context:                    # the 6 Ws — the underivable half
  who: distributed product teams
  what: real-time collaboration tool
  why: remove the latency tax of async handoffs
  where: AWS (ECS)
  when: production, since 2026
  how: TypeScript · React · Node · PostgreSQL

stack:                            # auto-detected
  frontend: react
  backend: express
  database: postgresql
  runtime: node
  deployment: aws-ecs
  build: vite
  testing: vitest
  cicd: github-actions

Validation rules: must have faf_version, project.name, project.goal · valid YAML · no secrets/credentials · works across all AI platforms.

The honesty rule

The engine only scores what's truthfully there — detection fills the stack, you fill the underivable Ws, and what nobody can source stays honestly empty. Empty beats wrong. That's why the resulting context is trustworthy.

Best practices — drive to Trophy

faf auto                    # 1. baseline: detect + seed
faf score                   # 2. see the empty slots
faf go                      # 3. close the gaps (guided)
faf check                   # 4. IANA compliance
faf sync                    # 5. emit CLAUDE.md / AGENTS.md
git add project.faf CLAUDE.md AGENTS.md && git commit -m "Add AI context (.faf)"

The .faf is the standard — commit it, and everyone's AI shares one context. Common boosters: add the human_context block, specify deployment, document the testing strategy.

Ecosystem

100k+ downloads across npm + PyPI + crates.io. claude-faf-mcp, faf-mcp, and gemini-faf-mcp hold AAA on Glama (earned, not conferred).

Standing

• IANA-registered media type: application/vnd.faf+yaml
• In the original Anthropic MCP ecosystem (PR #2759, merged Oct 2025)
• Maintainer of the .faf format specification

Troubleshooting

MCP server not responding

faf check                                   # is the .faf well-formed?
# re-check claude_desktop_config.json (see Quick start), then restart Claude
tail -f ~/Library/Logs/Claude/mcp-server-claude-faf-mcp.log

Low AI-readiness score

faf score          # shows the empty slots holding you back
faf go             # guided fill — close them to 100%

Tools not showing — restart Claude after editing the config, and confirm the server started (see the MCP log above).

Get this skill

• Plugin marketplace (recommended): /plugin marketplace add Wolfe-Jam/faf-skills → /plugin install faf@faf-skills
• Manual: git clone https://github.com/Wolfe-Jam/faf-skills.git && cp -r faf-skills/skills/* ~/.claude/skills/
• Browse the family: https://skills.faf.one

Resources

• Website: https://faf.one · Skills Site: https://skills.faf.one
• faf-cli: https://github.com/Wolfe-Jam/faf-cli · npm: https://npmjs.com/package/faf-cli
• claude-faf-mcp: https://npmjs.com/package/claude-faf-mcp

MIT · part of the FAF skill family (faf-context · faf-wizard · faf-expert). The format is the foundation — without .faf, there is no portable context.