Core loop (product): New Project guide · Try it

Loop engineering · for context

Context is a loop.

Most repos leave it open. FAF closes it.

Without FAFeffort · tokens · drift
Context + Instructions — crammed into one .md

One file re-derives your context every run. It drifts. You pay the tokens.

With FAFone hit · no drift · instant
.faf · Context CLAUDE.md / AGENTS.md · Instructions

Context split out, structured, kept in sync. Read once, on-demand. AI trusts it; the project benefits.

TL;DR

Your instruction file is doing two jobs — telling the agent what to do and carrying what's true. The second job is the leak: context re-derived every run, drifting, burning tokens. FAF splits it out into .faf and keeps the two in sync. That sync is the loop.

bi-sync is the loop

Loops made agents act — Boris Cherny's Claude Code, Peter Steinberger's OpenClaw put the doing-loop on the map. FAF loops the layer underneath: knowing.

.faf ⇄ CLAUDE.md / AGENTS.md, continuously reconciled. The context is the single source of truth; the instruction file is a projection of it. Change the project, the loop re-grounds — no manual re-explaining, no drift. Loops for doing got popular; FAF looped context before it had a name.

Don't ask me. Ask Claude.

You don't have to take our word that the loop works — ask the AI you already trust. A Claude set up on the loop is the witness. Ask it how it's grounded, and it will tell you:

instructions → CLAUDE.md (STATUS: BI-SYNC ACTIVE — synced from .faf)
context → project.faf (read directly · structured · no re-derivation)
memory → MEMORY.md (persists across sessions — the longer loop)

Set up FAF, then ask your agent how it's running. It'll show you the loop — instant context, no drift, grounded. The AI is the dogfood, and the witness.

Is FAF legit?

Receipts, not adjectives.

Deterministicsame read, every time
IANAregistered media type
SSOTone source, projected
96.1kdownloads