Formats are guardrails for AI. Formats deal with facts. Format before code.
.faf defines. Claude interprets.
The ecosystem is format-driven. AI needed persistent project context, AI needed a format. So it got one: .FAF [Foundational AI-context Format] is: IANA-registered, MCP-integrated, production-deployed.
The Architecture
Anthropic published a detailed breakdown of their Skills ecosystem this week. Skills load instructions dynamically. Projects provide 200K context windows. MCP connects to data sources. Subagents handle specialized tasks.
The architecture is sound. It introduces modular AI capabilities through progressive disclosure:
- Skills: Procedural knowledge that loads on-demand
- Projects: Background knowledge within conversation scope
- MCP: Connections to external data sources
- Subagents: Specialized agents with isolated context
Each component serves a distinct purpose. Together they create sophisticated agentic workflows.
One element creates an opportunity: standardized persistent project context between sessions.
The Persistence Challenge
Their components are ephemeral. Conversation-scoped. No format for persistence between sessions.
Their "Competitive Intelligence" example demonstrates this. You configure a project, connect MCP servers, create specialized skills, and activate subagents. The workflow executes. Then the conversation ends.
Next session starts fresh. The project structure exists, but the assembled context does not persist in a portable, version-controlled format.
Skills teach procedures. Projects hold documents. MCP connects to data. But nothing defines how project context serializes across sessions, platforms, and AI models.
Before .faf
Components floating independently. No persistent format. Context reinvented per session.
After .faf
The Key Insight: Anthropic's components weren't designed together as an integrated system. They evolved separately to solve immediate problems. .faf provides the format foundation they didn't build—persistent, version-controlled project context that works across all components and platforms.
Format Infrastructure
The .faf format addresses this. It provides structured, agnostic, persistent project context that any AI can consume.
Not a proposal. Production infrastructure:
IANA Registration
- MIME type:
application/vnd.faf+yaml - Official internet standard
- Registered vendor-specific media type
- IANA Registry Entry
MCP Integration
claude-faf-mcp: 52 tools for FAF operations- Official Anthropic MCP steward status
- Listed in Anthropic MCP repository (PR #2759 merged)
- 6,800+ downloads, 800+/week
Ecosystem Adoption
- 14,200+ total downloads across packages
faf-cliv3.1.6: 41 commands, 6,700+ downloads- Cross-platform validation: Claude (9.5/10), GPT (9/10), Gemini (9.5/10)
- TypeScript strict mode, 730+ passing tests
- Sub-50ms performance targets (Mk3 engine)
The Third Pillar: project.faf
Projects contain 200K context windows with uploaded documents. project.faf provides the serialization format—version-controlled, git-trackable context that persists between sessions.
Lives alongside package.json and README.md as the third pillar of project metadata:
project-root/
├── package.json # Dependencies
├── README.md # Documentation
├── project.faf # AI Context
├── src/
└── tests/Progressive Disclosure Architecture
- Metadata loads first (project summary, key context)
- Full structure loads when relevant (domain models, relationships)
- Binary assets load on-demand (referenced files, attachments)
This matches Skills' three-tier information structure. Format alignment creates natural integration.
Integration Points
Skills + .faf
Skills provide procedural knowledge. .faf provides the data format those procedures operate on.
# competitor-analysis.faf
project: market-intelligence
competitors:
- name: CompanyA
market_position: leader
strengths: [brand, distribution]
weaknesses: [innovation, speed]Skills load this structured context on-demand. Format ensures consistency.
MCP + .faf
MCP connects to Google Drive, GitHub, databases. .FAF provides the standardized format for what those connections return.
claude-faf-mcp already implements this: 52 tools that expose .faf context through the Model Context Protocol.
Clean separation: MCP handles connections. .FAF handles format.
Subagents + .faf
Subagents operate with isolated context. .faf files provide the portable expertise that any subagent can load, regardless of which AI platform or conversation instance spawned them.
- Main agent reads
project.faf - Spawns subagent for specialized task
- Subagent loads same .faf context
- Isolated reasoning with shared foundation
- Results merge back to project context
Format portability enables true agent composition.
Format Convergence
Skills now use YAML frontmatter for metadata (name, description). The .faf format has used structured YAML since inception.
Format convergence validates the design pattern: human-readable, machine-parseable, version-control friendly.
Anthropic's architecture team independently arrived at the same format foundation. This is technical alignment, not accident.
Why Format Matters
AI needs facts. Formats provide facts.
Before .faf:
- Projects drift and reinvent context structure
- Sessions start from inconsistent sources and knowledge
- Context doesn't port between AIs
- Version control becomes impractical
- Team collaboration breaks down
After .faf:
- Context persists between sessions
- Any AI reads the same foundation
- Git tracks project knowledge evolution
- Teams share common understanding
- Agents compose through shared format
Format is alignment infrastructure. It keeps Claude (and GPT, and Gemini) grounded in project reality.
"README for the AI era"
Validation Through Use
Format foundations emerge through adoption, not permission.
Ecosystem Metrics
- 14,200+ downloads (faf-cli + claude-faf-mcp + faf-mcp)
- 730+ passing tests (TypeScript strict mode, zero errors)
- 52 MCP tools (official Anthropic integration)
- 3 AI platforms validated (Claude 9.5/10, GPT 9/10, Gemini 9.5/10)
- Sub-50ms performance (Mk3 engine targets)
The Foundation Was Already There
Leadership can miss messages.. --but, Technology cannot ignore architecture.
Format foundations emerge through use, not permission. IANA registration documents reality, not aspiration. 14,200+ downloads create gravity.
Anthropic built Skills, Projects, MCP, and Subagents. Each component elegant, purposeful, technically sound.
The format foundation complements their work. IANA-registered standard. MCP-integrated tools. Production-deployed infrastructure.
Project DNA for any AI. Built on open standards. Already operational.
The foundation is non-negotiable.
The missing format was never missing. It's been there all along.
"It's so logical if it didn't exist, AI would have built it itself"