faf-wizard
Skill4 starsChanged 4 days agoMITView on GitHub
Detects the stack, reads what already exists (README, docs, an existing CLAUDE.md or .cursorrules), writes project.faf and scores it. The stack comes from the repo; the goal and six W''s are suggested from what exists, each with its source, and written only once the user confirms them. Use when someone wants project context set up for them: "set up Claude for this repo", "onboard AI to this codebase", "make me a project.faf". Works on new, legacy or inherited code. Turns the CLAUDE.md you already have into the structured facts behind it: GOOD to BETTER.
In this repository7 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- faf-contextSkill
- faf-expertSkill
- faf-goSkill
- repo-maintainerSkill
- wjttc-builderSkill
- wjttc-testerSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 4 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 61
Look for manifest files:
- Distinctive: line 69
No manifest? Still generate — ask user for stack info.
- Very specific: line 82
Create focused AI context at project root.
- Very specific: line 84
Write the stack from detection (faf init or faf auto does this, and leaves the goal and six W's empty).
- Distinctive: line 114
Keep under 200 lines. Only what changes AI behavior.
- Distinctive: line 167
When CLAUDE.md or .cursorrules exist:
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
<!-- faf:start -->v: 1
needs:
- "faf-cli on Node 22+ for the score (not installed: run `npx faf-cli score`)"
reads: [README.md, docs/, "CLAUDE.md or .cursorrules", package manifests, the directory structure]
writes: ["project.faf at the project root (stack from the repo; the goal and six W's only as the user confirms them)"]
network: ["registry.npmjs.org (read, only when npx fetches faf-cli)"]
publishes: false
done_when: "project.faf exists at the project root and has a score"
project: project.faf<!-- faf:end -->FAF Wizard
The pit crew. Point it at any project, get a scored
.faffile back: the repo's facts filled in, the goal and six W's suggested for you to confirm. See also:faf-expert— the mechanic's manual for when you want to understand every bolt.
What This Does
Generates a project.faf file for any codebase. You don't need to know the format: the wizard handles detection, generation, scoring and reporting. The one thing it never does is decide your project's goal and six W's for you: it suggests them, and you confirm each one.
You: "Set up project context for this repo"
Wizard: Detects stack → writes project.faf → suggests the goal and 6 W's → you confirm → scores it → reports gapsThe Problem It Solves
Even famous repos have no AI-context. React. Kubernetes. Rails. Django — none ship a .faf. Detection alone gets React to ~56%; the human half (who / why / where) that grounds an AI is missing. That's the gap.
| What exists | What it answers |
| README.md | "What does this do?" (for humans) |
| docs/ | "How do I use this?" (for humans) |
| project.faf | "How should AI help build this?" (for AI) |
Documentation tells humans how to use code. AI-context tells AI how to help you write it. They're not the same thing.
Works on ANY Project
| Project State | What FAF Wizard Does |
| Brand new | Optimized AI context from line one |
| Legacy codebase | AI finally understands the archaeology |
| Famous OSS | Even React needs this (it doesn't have one) |
| Side project | Stop re-explaining every session |
| Client handoff | Portable context for any AI tool |
Workflow
Step 1: Detect Stack
Look for manifest files:
package.json→ Node.js/TypeScriptCargo.toml→ Rustpyproject.toml→ Pythongo.mod→ Gopom.xml→ JavaGemfile→ Ruby
No manifest? Still generate — ask user for stack info.
Step 2: Mine Existing Context
Pull signals from what exists. These are sources for suggestions, never values written straight into the goal or the six W's:
- README.md (the first paragraph, a
## Whyor## Whatsection) - docs/ (architecture decisions)
- CLAUDE.md or .cursorrules (existing AI context, to migrate)
- Package manifests (the
description; dependencies reveal the stack) - Directory structure (architecture patterns)
Step 3: Generate project.faf
Create focused AI context at project root. Every .faf is scored against the same 33 slots: faf-cli fills the 21 base slots and marks the 12 enterprise slots slotignored unless your app type uses them. app_type selects which are active; the rest are slotignored (never counted against you).
Write the stack from detection (faf init or faf auto does this, and leaves the goal and six W's empty). Then go to Step 3b: the goal and six W's are the person's to give.
faf_version: "3.0"
project:
name: my-project
goal: A task API that tells small teams who owns what
main_language: typescript
human_context: # the user's words, confirmed in Step 3b
who: Small teams tracking shared tasks
what: A REST API for task boards and assignments
why: Spreadsheets lose who owns what
where: Vercel (serverless), npm
when: Beta since March 2026
how: Express REST handlers over PostgreSQL; a queue worker sends reminders
stack:
frontend: react
backend: express
api_type: rest
runtime: node
database: postgresql
hosting: vercel
build: vite
testing: vitest # informational — not one of the 33 scored slots
cicd: github-actionsKeep under 200 lines. Only what changes AI behavior.
Step 3b: Suggest the goal and six W's, the user confirms
The goal and the six W's are human slots: AI suggests them, a person confirms each one. Never write one the user hasn't seen and approved.
- Draft a suggestion only where something exists to quote, with its source: the README's first paragraph for the goal, the package
descriptionforwhat, a## Whysection forwhy, and an existing CLAUDE.md## Contextblock (CLAUDE.md § Context) for any W it states. Words a person already wrote are still confirmed, not copied. Never offer an install line, a badge or link row, a notice ("deprecated", "sponsored by"), or a heading. - Ask with AskUserQuestion, one question per slot: the suggestion and its source (
README.md § Why), with use it · edit it (Other) · skip. No suggestion: ask the question, with the slot's definition as help. - Write only what the user confirmed, word for word:
faf edit human_context.why "<their text>"(orproject.goal). Skipped slots stay empty.
What each slot holds (one line, every word earns its place):
- goal: what it does, and for whom
- who: the people who use it, not the author
- what: what they make or do with it
- why: the problem it removes
- where: where it runs or ships
- when: the stage it's at, and since when (not a version or a date to ship)
- how: how it works: the mechanism, and the choice that shapes the code (not an install command or a list of the stack)
Step 4: Score and Report
Generated: project.faf
AI-Readiness: 87% ◇ Bronze — Production ready
Filled: 13/15 active slots
slotignored: the slots that don't apply to this app_type (never counted)
To reach 100% ✪:
- Add deployment details
- Document the testing strategy
- Confirm the goal and W's still empty (faf-go asks for them)The target is 100% ✪ — Bronze is the floor, not the finish line. At 100% every active slot is filled and the AI starts every session already knowing the project.
Tier System
| Score | Tier | Symbol | Status |
| 100% | Trophy | ✪ | AI is optimized to code |
| 99% | Gold | ★ | Exceptional |
| 95% | Silver | ◆ | Top tier |
| 85% | Bronze | ◇ | Production ready |
| 70% | Green | ● | Solid foundation |
| 55% | Yellow | ● | Needs improvement |
| 1% | Red | ○ | Major work needed |
| 0% | White | ♡ | Empty |
Target: 100% ✪. Bronze (◇ 85%) is the floor where AI collaboration becomes productive — Trophy is the goal: every active slot filled, the AI fully grounded from session one.
Migration
When CLAUDE.md or .cursorrules exist:
- Offer to migrate to .faf (portable format). Migration moves the stack and the rules; it never writes the goal or a W directly. Those go through Step 3b as suggestions (
CLAUDE.md § Context), each confirmed by the user - .faf works in Cursor, Gemini, WARP, Windsurf — not just Claude
- YAML structure parses better than freeform markdown
Validation
Generated files must:
- Hold no goal or W the user didn't confirm
- Be valid YAML
- Have
faf_versionandproject.name - Contain NO secrets or credentials
- Stay under 200 lines
Credentials
- IANA Media Type:
application/vnd.faf+yaml - Listed in modelcontextprotocol/servers (#2759, merged 2025-10-17)
- Website: https://faf.one
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

