How to Build an Agentic OS with Fable 5 – The Builder’s Guide

Fable 5 is the most capable model available right now. Used casually, it is an expensive way to generate impressive mistakes. Used inside a proper system, it becomes the closest thing to a full employee you can rent for a few dollars a day.

This guide gives you the actual files, order, and checkpoints to build that system. Not theory. Not philosophy. Concrete structure that runs unattended, stays honest, and improves over time.

What you end up with after 2 hours of setup:

• A rock-solid CLAUDE.md constitution the model cannot negotiate with

• A daily heartbeat loop where Fable 5 makes decisions but spends almost no tokens (cheap models do the work)

• An independent verifier that never grades its own homework

• A trust ledger that automatically grants and revokes autonomy per skill based on real pass rates

• A goals directory where finished work becomes re-verified invariants that never rot silently

• A self-enforcing budget that warns you before you overspend

• Optional advanced loops (quorum, ratchet, sparring, compost) with clear install conditions

Prerequisites

• Bash + terminal

• Claude Code with Fable 5 access (usage credits enabled)

• llm CLI with OpenRouter key for cheap models

• jq, git, make, cron

• A repo with at least one test command

The 3 core principles (everything follows these)

1. Laws, not tips. Every rule has a number, a "never", or a command that checks it. Soft advice gets optimized away.

2. Nothing grades its own homework. Planner, worker, verifier, and gate are separate parties. The last one is deterministic.

3. Finished work graduates into invariants. A goal you verify only once is just an assumption with a timestamp.

BUILD 0: Engine Configuration (from Anthropic’s own docs)

Use these official settings:

BUILD 1: The Constitution (CLAUDE.md)

This file is non-negotiable. Put it at the root of your repo.

Key rules it must contain:

• Never exceed X changed lines without asking

• Never touch auth, billing, migrations, or prod config unattended

• Done = the verify script passes. Nothing else counts.

• Never invent secrets, endpoints, or conventions

• Never add dependencies without proposal in STATE.md

• Never report work as done from self-assessment

Keep it under 150 lines. Every line should have a number, a never, or a verifiable command.

BUILD 2: Walls and Gate

Create contract.md with three sections: acts alone, queues for me, wakes me up.

This script must exit 0 on clean code. The whole system stands on it.

BUILD 3: The Heartbeat (loop.sh)

This is the core loop that runs daily via cron.

Triage stage (cheap model) decides if anything is actionable.

Conductor (Fable 5) decides the next best action.

Worker executes.

Verifier checks.

Gate decides ship or review.

All communication happens through structured work orders. Models only see what they need.

BUILD 4: Trust Ledger

Autonomy is per skill, not per agent. Track runs and pass rates in trust.tsv.

• 20+ runs and 95%+ pass rate → auto ship

• Below thresholds → watch or queue

Demotion is automatic and loud. The ledger tells the truth about every skill.

BUILD 5: Standing Goals

Finished work graduates into goals/ folder. Each goal has a predicate (shell command) that is checked daily.

If a goal ever fails, it wakes you and lists suspects. Finished work never rots silently again.

BUILD 6: Budget Enforcement

Log every spend. Daily check stops the loop if you hit the cap. Know exactly what each stage costs.

BUILD 7-8: Advanced Loops & Ops

Add quorum voting, ratchet metrics, sparring (builder vs breaker), weekly compost (propose system improvements).

30-Day Trust Schedule

Week 1: supervised, manual ticks

Week 2: drafts with review

Week 3: auto ship on trusted skills

Week 4: grow, prune, and let the system improve itself

After 30 days you have a real Agentic OS that runs mostly unattended, stays honest, and compounds capability over time.

This is not another prompt collection. This is the actual system files and rules that turn Fable 5 from impressive demo into reliable infrastructure.

Start with BUILD 1 tonight. Do the checkpoints. The first time verify-goals.sh catches a silent regression on something you thought was finished, you will understand why this structure matters.

The model was never the hard part. Building something around it that stays honest when you stop watching - that was the hard part.

You just built it.

You build your own life.

Prompts

- Model: claude-fable-5
- Context: 1M tokens, up to 128k output
- Pricing: $10/M in, $50/M out
- Effort: high (default) for most work, xhigh only for long conductor reviews
- Never use prompts that ask it to “show your thinking” — this triggers refusals.

Create guardrails/verify.sh - your deterministic final gate (npm test, typecheck, lint, etc.).
 

Makefile + cron ties everything together.
 

Links

• loop.sh
• verify-goals.sh