Anti-hallucination protocol
Every important claim must carry an evidence tag. Nothing gets asserted about your project without inspection.
A copy-paste operating system for Claude Code
Two blocks of text that turn Claude Code into a full command structure — an overseer, a manager, five-agent councils, builder crews, watchdogs, heartbeats, and a completion gate that only opens on verified evidence. No hallucinated files. No vanished tasks. No fake autonomy. No flattery.
Copy Part 1 into Claude's settings (general / personal preferences) — or the top of your global ~/.claude/CLAUDE.md. Every conversation runs truth-first from then on.
Open Claude Code inside the project you want governed and paste Part 2 — all 44 directives — as one single message.
It maps your repo before touching anything, builds the agent system the router points at, then proves it works by running a live test task through the whole chain.
Chain of command
What it installs
Every important claim must carry an evidence tag. Nothing gets asserted about your project without inspection.
Architect, engineer, security, product, and an adversary who assumes the plan is wrong — each reasons independently before synthesis.
The overseer is monitored by an independent mechanism. Liveness is proven by heartbeats, never by an agent's own claim.
Every request gets a task ID and an explicit state machine. Crashes resume from persistent state — work is never silently restarted.
"Done" only after tests actually ran and the result was validated against the original request — not merely whether it compiles.
A schedule only exists if a real scheduler is configured — cron, Task Scheduler, a daemon. Agents never pretend they'll wake up later.
Average Claude vs. this
The honest part: almost nobody runs Claude like this. Third-party estimates put Claude at a few hundred million users — and Claude Code, the tool this system runs in, at roughly four million weekly developers. That's already a small fraction, and most of it runs stock defaults. A persistent, supervised, self-recovering setup is a fraction of that fraction. This takes someone using Claude casually and hands them a setup a small percentage of the world runs — and installing it is literally a copy-paste. The gap between casual and this isn't talent. It's knowing the system exists.
Together, the three pieces on this page — the agentic operating system, the truth-first general instructions, and a persistent memory like the Obsidian vault — are the same three pillars serious, enterprise-grade AI systems are built on: orchestration with governance, an explicit policy layer, and durable knowledge. That combination is the foundation. Everything after it is tuning.
Make it permanent — attach a memory
A chat's memory dies with the session. Files don't. The single biggest upgrade to this system is connecting Claude to a persistent, file-based memory so its rules, decisions, task ledgers, and lessons are reloaded at the start of every session — automatically, in every project, for as long as the files exist.
Point Claude at a local Obsidian vault (via MCP or its Local REST API) and have it read a master note first, then write decisions, research, and session logs back. Your knowledge base becomes the AI's long-term memory — and it compounds.
The zero-dependency version: a global CLAUDE.md plus a memory directory of small fact files with an index. Claude Code loads them every session. The mega-prompt's task ledger and decision records live the same way.
For large bodies of knowledge, add retrieval: embed your docs and notes so the relevant pieces are pulled into context on demand. Heavier to set up, strongest at scale.
Install Obsidian and create a vault. Under the hood it's just a folder of Markdown files on disk — which is exactly why an AI can read and write it.
Add one note at the vault root (name it something like _CLAUDE.md) that tells the AI how the vault is organized: what to read first, where daily notes live, where decisions and research get filed. Every session starts by reading this note.
Claude Code: no plugins needed — give it the vault folder's path in your global CLAUDE.md with the rule "read the vault's _CLAUDE.md before touching anything." Claude Desktop / claude.ai: connect a filesystem MCP server pointed at the vault folder, or install Obsidian's community Local REST API plugin plus an Obsidian MCP server on top of it (that route only works while Obsidian is open).
The rule that makes it compound: every working session ends with Claude saving decisions, research, and lessons into the vault. Reading the vault makes the AI informed. Writing back makes it permanent — the memory survives every new chat, forever, as long as the files exist.
Or skip the manual steps entirely — copy the prompt below into Claude Code and it does the whole setup for you, asking for your input only when a step truly needs you.
Before you run it
Installed Claude Code required. The install — reading files, creating the vault, wiring the memory — has to run in the installed version of Claude Code (terminal or desktop app) on a Windows / Mac / Linux computer. Not the browser version, and not a phone. Once it's built you can sync your systems across devices, but the setup itself happens on your machine. Read this page anywhere; save the link and run the install from your computer.
Run the bootstrap on Fable 5 / 5.1 — the strongest agentic models for a build like this. Fable is included on Max plans; on Pro it runs on pay-as-you-go usage credits. No Max and no credit budget? Use Opus 5 and turn the effort up.
This is a real build — inspection, councils, implementation, failure simulation, audit — not an instant answer. Without a Max plan, expect it to take two to three full sessions of credits. That's fine: the system resumes unfinished work, it never restarts from zero.
Turn the model's effort all the way up — run the bootstrap with Ultracode on. Ideally you're already on Fable 5 / 5.1; if you're on Opus 5 without a Max plan, Ultracode effort matters even more. Smarter is better when you're setting up the system that governs everything else.
Part 1 — The general instructions
Paste this into Claude's settings (Settings → Profile → personal preferences on claude.ai, or the top of your global ~/.claude/CLAUDE.md for Claude Code). It makes every conversation truth-first: evidence over agreement, uncertainty labeled, zero flattery. Swap [YOUR NAME] for your real name — if Claude ever stops using it, that's your early-warning signal the instructions have drifted out of context.
Part 2 — The bootstrap mega-prompt
Paste this into Claude Code as one message, inside the project you want it to govern. It inspects first, builds second, and refuses to call itself finished until a live test task survives the full chain of command.
Why it's free
The system was never the advantage.
This is given away on purpose. A framework doesn't get things done — the person prompting it does. The power is in the operator: the questions you ask, the standards you refuse to lower, the follow-through when validation fails. Handing you the blueprint costs the builder nothing and changes nothing. So take it, install it, and run it hard. What you build under it is yours.
Was this helpful?
If this page gave you something, say so — it's the only payment asked. One click, counted anonymously.
Logged. Signal received — thank you.