When the Frontier Model Visits, Make It Rebuild Your Setup

Access to Claude Fable 5, Anthropic's frontier model, comes and goes. My first window lasted three days a few weeks back before access closed. Yesterday it opened again. I have no idea how long this one lasts, and when frontier models move to metered API pricing, every token it thinks with will show up on a bill.

The temptation with a window like that is obvious: cram in as much work as possible before it closes, then go back to normal and miss it.

I did something different. I gave it a single instruction: you will not be here long, so build the things that will make the regular model run with your mind after you are gone.

What it built

It studied its own behavior and turned the patterns into files. Not notes. Not a document of tips. Actual harness artifacts that load and enforce themselves in every future session:

• A doctrine skill that auto-fires on any substantive task and walks the model through the loop a frontier model runs by instinct: scout the real code first, commit to one plan, act to completion, verify like a skeptic, report outcome-first.
• A guard hook that pattern-matches my prompts and injects the doctrine straight into context. The model cannot forget what the harness keeps putting in front of it.
• A skeptic agent whose only job is to attack conclusions before they reach me. It returns CONFIRMED, REFUTED, or UNPROVEN, and it defaults to UNPROVEN unless it observed the behavior itself.
• Three multi-agent workflows: a deep-audit that sweeps a codebase with six parallel lenses and loops until two consecutive rounds find nothing new, a ship-gate that returns GO or NO-GO before anything customer-facing deploys, and a judge-panel that argues big decisions from genuinely distinct positions before recommending one.

Then it tested its own work. It ran the hook against matching and non-matching prompts, parse-checked the workflow scripts the way the harness executes them, and wired the whole kit into my project instructions so it loads on every session regardless of which model is running.

The insight that makes this work

Here is the thing I did not expect it to tell me: most of the gap between a frontier model and a smaller one is not intelligence. It is discipline.

The expensive model verifies before it claims. It exercises the change instead of reading the code and declaring victory. It keeps sweeping an audit after the first convenient handful of findings. It attacks its own conclusions. A smaller model is capable of every one of those behaviors. It just does not do them by default, the same way a talented junior does not run a preflight checklist by default.

Checklists are how aviation made average pilots perform like great ones. Skills, hooks, and workflows are the same move for AI agents. The discipline stops living in the model and starts living in the system, and systems do not have off days.

The economics, because they matter now

This is not just about quality. It is about the bill that is coming.

When frontier usage is metered, the wasteful pattern is paying frontier prices for work a cheaper model could do with the right harness. The kit flips that: the cheap model runs the frontier playbook on everyday work, and you save the expensive tokens for the problems that genuinely need them. Frontier discipline everywhere, frontier prices only where they earn it.

And if your quality depends on which model you happen to have this month, you do not have an operation, you have a subscription. Pricing changes, tiers change, access windows close without warning (mine did). The move is to treat every session with a stronger model as a chance to upgrade the system itself. Ask it to encode what it does differently. Make it write the skill, the hook, the workflow, the checklist. A three-day window becomes a permanent upgrade to every session that comes after it, on any model, at any price.

That is the actual work we do at Modern Mustard Seed, by the way. Not "AI tips." Operations where the intelligence is built into the system: voice agents answering phones at 2am, audit engines that find revenue leaks, agentic setups that run a business's back office. The models will keep changing. The systems compound.

Steal the whole thing

I published every artifact as a free playbook: the doctrine skill, the guard hook, the skeptic agent, the full deep-audit workflow with the adversarial verification panel, and the recipes for the other two. Copy-paste, ten minutes, no dependencies.

Get the Fable Mind Playbook, free.

And if you would rather have this installed, tuned to your stack, and pointed at your revenue: book a call. This is what we build.