Tell it what you want
and how long it has.
Brainfile puts Claude Code on Autopilot: it does the work, proves it, and gets better. A run protocol plus the files that make it stick — a backlog that never runs empty, a ledger that never forgets a directive until it's proven done, memory that compounds instead of resetting to zero, and a brief waiting for you at the end of every run.
Six features.
Every one ships as real files.
No hosted runner, no app to log into. Autopilot is a run protocol your own Claude executes in your own project — plus the plain files that make it compound. Everything below is in the pull on day one, including the bf launcher: type bf 8h in any terminal and a timed Autopilot run starts, ending with a written brief.
It selects the work, does it, and proves it.
Pre-flight reads your ledger + backlog and picks the highest-leverage plan for the time you gave it. The work loop verifies every item — "it exists" is not "it works". Post-flight re-checks the evidence and writes learnings back into your rules, so the next run starts smarter. Recurring schedules included — with the honest keep-your-computer-awake setup (you start each run; Claude can't open your laptop).
AUTOPILOT.md — the full protocol.Runs never stop early because the list ran out.
The backlog is built from work classes that refill themselves: each one states the next increment and the rule that generates the one after it. When a run plan is exhausted, Autopilot pulls the next highest-leverage increment instead of stopping.
brain/backlog.md — 5 starter classes, seeded with YOUR goals at interview.Every directive tracked until it's PROVEN done.
Anything you ask for gets a row that stays OPEN until there's an evidence line — what was checked and what it showed. "I did it" is not evidence. Items blocked on you surface in every brief; nothing is silently dropped.
brain/ledger.md — re-read at the start of every run.Questions that GENERATE your files.
Your first session is an interview: business, goals, rules, tools, your Autopilot schedule, your brief preference. Claude writes the answers into CLAUDE.md, seeds your backlog from your month-1 goals, and starts your decision log — so session 50 starts as warm as session 1.
INTERVIEW-ME.md · see the actual question set →Accomplished / needs you / next priorities.
Every run can end with a written brief: what shipped (with evidence), what's blocked on you, and a ranked recommendation for the next run. It's a plain file — your run dashboard without an app to log into. The template is yours to customize, and it's opt-in at interview time.
templates/END-OF-RUN-BRIEF.md — see a real example below.Decisions stop dying in scrollback.
An append-only log Claude maintains at the end of every session: what you asked, what was decided, what was deferred and where it's tracked. Auditable, searchable, and honest — corrections get a new entry, past entries are never rewritten.
CONVERSATION-HISTORY.md — the pattern + entry format.your-project/ ├── CLAUDE.md ← generated from YOUR interview ├── AUTOPILOT.md ← the run protocol ├── INTERVIEW-ME.md ← the questions that generate your files ├── START-HERE.md ├── CONVERSATION-HISTORY.md ├── templates/ │ └── END-OF-RUN-BRIEF.md └── brain/ ├── backlog.md ← never runs empty ├── ledger.md ← never forgets ├── decisions.md ← created in your first session ├── runs/ ← plans + briefs, one per run └── freshness/ ← the weekly living feed
# brain/ledger.md — the never-forgets ledger Every directive you give Claude gets a row here, and it stays OPEN until it is **proven** done — with evidence, not vibes. ## Rules for Claude 1. **Capture:** a directive not in the ledger doesn't exist. 2. **Prove:** PROVEN_DONE requires an evidence line — what was checked and what it showed. "I did it" is not evidence. 3. **Surface:** rows blocked on you get NEEDS_YOU and appear in every brief until resolved. 4. **Review:** every pre-flight lists OPEN rows, oldest first.
# Run brief — 2026-07-11 (one working session) ## Goal Rebuild brainfile.io around Autopilot — six v1 features, honest counts, new onboarding page. ## Accomplished (with evidence) - Home page reworked around the Autopilot story — evidence: this page; every feature above maps to a file in the pull - Pack manifest updated to honest counts (10 pattern files + 15 role packs) — evidence: call list_pack_contents on our live MCP server and count - Onboarding page with the real interview question set — evidence: /onboarding.html ## Needs you - Nothing — all clear. ## Next priorities (recommendation for the next run) 1. Collect votes from the "what should Autopilot learn next?" ladder below and build the winner. ## Learnings written this run - Rule added: every new page section ships with a feedback link.
The brief above is not a mockup — it's the end-of-run brief this run produced, in the same template your pack ships. Disagree with the feature set? Tell us →
Runs on your existing Claude subscription.
Other AI tools charge API costs on top of a subscription. Brainfile runs on YOUR Claude Code account — Pro, Max, or Team. No additional API charges. Ever.
The math is simple: Brainfile is the operating system. Claude is the engine. You already have the engine.
The gap between "free CLAUDE.md tips"
and a system that runs
Karpathy's CLAUDE.md is 66k stars and 1 file. Brainfile is the Autopilot protocol, 10 pattern files, 15 role packs, and a versioned update feed. Here's exactly what you get — browse every file in the box.
- —1 instruction file (hand-maintained)
- —No memory across sessions
- —No autonomous agents
- —No quality enforcement
- —Breaks on model updates
- —8+ hours to set up properly
- —Manual updates forever
- ✓The Autopilot protocol — tell Claude what you want and how long it has; pre-flight, verified work loop, post-flight learnings
- ✓Persistent memory — Claude knows your business across every session
- ✓10 pattern files + 15 role packs — never-empty backlog, never-forgets ledger, interview onboarding, end-of-run brief, conversation history, plus a role-tuned CLAUDE.md per profession
- ✓Interview onboarding — questions that generate YOUR files, not placeholders you fill by hand
- ✓Freshness feed — versioned update packs as Claude changes, every item traceable to a real change
- ✓15 min install — one pull command, let Claude interview you
- ✓Continuous improvement — new packs added as they prove out on our own businesses
30-day money-back guarantee · cancel anytime
The actual files in the box —
not a mockup.
Every excerpt below is unedited, straight from the pack you pull on day one. No fabricated screenshots, no fake reviews — just the real artifact, including the freshness feed that keeps it alive.
You re-explain your stack, your rules, and what you're building at the start of every session. Claude opens cold and generic each time — no memory of yesterday's decisions.
Claude reads your Brainfile — business, rules, decisions, and the latest freshness feed — and starts warm. It already knows your context on session 50 the way it did on session 1.
# Interview Me — run this first Paste this whole file's instruction into Claude Code in your project, or just say: > "Read INTERVIEW-ME.md and interview me to fill in my > Brainfile." ## Instructions to the agent You are setting up this user's Brainfile (their Claude Code operating system). Your job in this first session is to **interview the user and personalize their operating profile** so every future session starts already knowing their business. Do this now: 1. **Ask, don't assume.** Interview the user with these questions — ONE at a time, waiting for each answer: - What's your business/project, in one line? Who does it serve? - What are you trying to get done with Claude over the next month? - What are your top 3 non-negotiable rules (things Claude must always or never do)?
# CLAUDE.md — Starter Operating System (Brainfile)
This is a starter operating profile. Replace the bracketed
parts with your business, then let the freshness feed
evolve it.
## Who you are working for
- Business: [your business]
- What we do: [one line]
- Who we serve: [customer]
## Operating rules (edit freely)
1. **Think before act.** Clarify the ask, map where it
belongs, check current state, plan, build, verify.
2. **Spec before build.** No feature without a 3-line
written spec.
3. **Verify before "done".** "It runs" is not "it's
correct" — check the actual output a user sees.
4. **Every error becomes a permanent check**, not just a note.
5. **Keep a memory.** Persist decisions in `brain/` so
sessions don't reset to zero.
# Brainfile Freshness Feed — v2 (2026-07-04)
Your weekly Brainfile update. Recurring value: the same improvements we ship to our own
operating system, plus what changed in Claude Code — with how to apply each to your brain/.
(Dogfood: we run these on our own six live businesses.)
## What we shipped to our own Brainfile (apply to yours)
- **START-HERE.md orientation in every role pack.** Activation guidance now travels INSIDE
the artifact. *Apply:* add a root START-HERE.md so cold-opening the folder orients you in
30 seconds (install -> confirm-it-loaded prompt -> first personalizations -> how it compounds).
## What changed in Claude Code (real release highlights, source-linked)
Source: https://code.claude.com/docs/en/changelog
- Changed the "default" permission mode to "Manual" across the CLI, VS Code, and JetBrains.
- `AskUserQuestion` dialogs no longer auto-continue by default; opt into an idle timeout via `/config`.
- Improved screen-reader output: decorative glyphs hidden, transcript symbols read as short labels.
This is the real v2 feed shipping in the pack today. The next feed ships with the next Claude Code change — we never backfill fake updates.
Don't take our word for it.
Poke the live infrastructure yourself.
Everything below is running in production right now — an open MCP server, a public registry listing, machine-readable docs. Every claim on this page is verifiable from your own terminal before you spend a dollar.
$ curl -s https://brainfile.io/mcp {"ok":true,"server":"brainfile","version":"1.0.0", "transport":"streamable-http (JSON-RPC 2.0 over POST)", "tools":["get_starter_file","list_pack_contents","get_freshness_sample"], "scope":"free (no token) — the on-ramp"}
get_starter_file
Returns the full, unedited text of the three starter files — INTERVIEW-ME.md, START-HERE.md, and CLAUDE.starter.md. The same files the paid pull ships. No account needed.
list_pack_contents
Returns the whole pack manifest: the 10 free pattern files (5 readable right here over MCP, all 10 in the free no-card trial pull), plus the 15 role packs — listed honestly as locked, with the unlock link.
get_freshness_sample
Returns the actual freshness-feed entry shipping in the pack today (markdown + structured feed.json) — the weekly living-system update that keeps your install current.
Point Claude, Cursor, or ChatGPT at https://brainfile.io/mcp and call the tools yourself — 2-minute setup guide →
Listed in the Official MCP Registry
Brainfile is published in the Model Context Protocol's official registry as io.brainfile/brainfile — domain-verified, status: active. Any MCP client can discover and connect to it.
io.brainfile — see the live listing →Machine-readable by design: llms.txt
We practice what we sell. This site publishes /llms.txt — a plain-text summary AI assistants read directly, cross-linked with our other properties — so when someone asks their AI about Brainfile, it gets the honest source, not a scrape.
The "interview me" onboarding
INTERVIEW-ME.md is a markdown guide in the pack: open Claude Code and say "Read INTERVIEW-ME.md and interview me to fill in my Brainfile." Claude asks about your business, goals, and non-negotiable rules — one question at a time — then writes the answers into your CLAUDE.md and starts your decision log. It's instructions your own Claude follows, not an app we run.
get_starter_file — or read the excerpt above.Five patterns that separate
"an AI with a prompt" from an operating system.
These are the working patterns we run every day on our own six live businesses. Your pack ships each pattern as plain files and rules your Claude reads — the pattern, not our internal systems.
Persistent memory files
A CLAUDE.md + brain/ folder your AI reads at the start of every session. No re-explaining your stack, rules, or goals — Claude knows your business on session 50 the way it did on session 1.
Nothing is "done" without evidence
The rule we live by: work isn't complete because the AI says so — it's complete when the output is verified. On our own system, tasks literally cannot be marked done until an evidence file exists. Your starter ships the same rule: "Verify before 'done' — 'it runs' is not 'it's correct'."
Self-improving checklists
Every error becomes a permanent rule, not a note that gets forgotten. Correct your AI once and the correction is encoded — your operating system gets stricter exactly where you got burned.
Registries — nothing gets forgotten
Anything worth maintaining gets a registry entry with a review cadence. If it's not registered, that's a blind spot to fix. It's how a growing AI setup avoids the silent rot of things nobody is watching.
Decision logs
Decisions get written to brain/decisions.md once — and never re-litigated. Your AI stops asking you the same questions and stops quietly reversing choices you already made.
Get the Brainfile OS Starter Kit — Free
See what a production-grade CLAUDE.md looks like before you buy:
- ✓Annotated CLAUDE.md — every section explained
- ✓Sample brain/ memory structure for a solo founder
- ✓1 pre-built agent (Revenue Agent) — plug in and run
No spam. We send a monthly update on what changed in Claude + your install. Unsubscribe anytime.
One plan to start. Everything in v1.
The Autopilot protocol, all 10 pattern files, all 15 role packs, and the freshness feed — one flat price, on your existing Claude plan.
Not ready to commit? Get the $249 AI-Ops Audit first.
We point the same drift detector, deploy gate, visual-QA, and launch-readiness gates that guard our own six live businesses at your Claude Code setup — and hand you a scored report + prioritized fix list. One-time, no subscription. It's the low-friction way to see exactly where your setup stands before you run the full OS.
30-day free trial first — no card. Cancel anytime.
- The Autopilot protocol — tell Claude what you want and how long it has; it does the work, proves it, and gets better
- All 10 v1 pattern files — never-empty backlog, never-forgets ledger, interview onboarding, end-of-run brief template, conversation history, memory starter
- All 15 role packs — a role-tuned CLAUDE.md per profession (Founder, Developer, Trader, Consultant, Lawyer…). Subscriber-only — the free trial ships everything else, not these.
- Freshness feed — versioned update packs as Claude Code and best practices change, traceable to real changes (never filler)
- Works on your existing Claude plan — Pro, Max, or Team; $0 additional API costs
- 30-day money-back guarantee — cancel anytime
Prefer to pay yearly? $999/year — the same single plan, billed annually ($83/mo). See what's included →
What should Autopilot learn next?
v1 is the six features above. The next capability gets built when it wins the vote — one click, no email needed. We don't display counts (no fake momentum); the ladder is real and we build the winner.
Something else entirely? Tell us what Autopilot should learn →
From zero to your first Autopilot run
in 15 minutes.
No coding experience required. Install is one pull command. Configuration is a first-session interview Claude runs for you.
📦 Install
5 minutesStart your trial and we email you a one-line pull command (also shown on screen). Run it in your project root to pull Brainfile in — re-run it anytime to get the newest freshness feed: curl -sSL https://brainfile.io/pull.sh | bash -s -- [email protected] <token>. That's the install.
⚙️ Configure
10 minutesStart claude and say "interview me and fill in my Brainfile." Claude walks the [CUSTOMIZE] markers — your business, goals, metrics, preferences — and writes them in. No generic setup.
🚀 Run
Ongoing — compoundsSay "Read AUTOPILOT.md. You have 2 hours. Run autopilot." Claude selects the work, does it, proves it, and leaves you a brief. Memory and rules compound run by run.
Common questions.
caffeinate / Windows power settings) and a recurring schedule recorded in your CLAUDE.md at interview time. Nothing runs on our servers; your work never leaves your machine. Ready to put Claude Code
on Autopilot?
30-day free trial, no card. Works on your existing Claude subscription.
30-day money-back guarantee · Cancel anytime · [email protected]