BLITZMETRICS · META / HOW WE DO IT

We told Claude to pick its own model — here’s what happened

A one-line tip from Simon Willison became a standing rule across our whole operation. Let the AI decide which engine each task needs. Cheap for the grunt work, the ceiling for the judgment calls. This is exactly how we wired it.

By Dennis Yu · BlitzMetrics · Built with a Claude agent, July 6, 2026

Simon Willison posted a note on July 3 that stuck with me.

The tip came from Cat Wu and Thariq Shihipar on the Claude Code team: don’t dictate to a smart model how it should work. Let it use its own judgment. Their example was tests — instead of writing a rule like “only test big features,” just tell the model to decide when a test is worth writing.

Then Jesse Vincent added the part that made me sit up. Tell the model to use cheaper models for the small stuff — and let it decide which one. Simon typed one sentence into Claude Code: “For all coding tasks, use your judgment to decide an appropriate lower-power model and run that in a subagent.” His Fable budget started lasting a lot longer, and he got more done, not less.

My first thought: is this a coding thing? Because we don’t write code all day. We score personal brands, harvest mentions, publish across 198 sites, write in people’s voices. Does “let the model pick the model” apply to us?

It does. More than it applies to code, actually. So we made it a standing rule, put it in the skill packs we hand out, and shipped it into the audit engine. Here’s the whole thing.

There are two versions of this, and they’re mirror images

Once I dug in, it turned out there are two patterns, a month apart, that people keep confusing. They’re opposites — and you want both.

Delegate down is the default. The smart model does the thinking, then spins up cheaper subagents for the high-volume grunt work and reviews what comes back.

Escalate up is Anthropic’s Advisor Strategy from April. Flip it around: a cheap model like Haiku drives the whole task and only pings Opus when it hits something it can’t handle. Anthropic’s own numbers on this are not subtle — Haiku’s score on hard web research went from 19.7% to 41.2% with an Opus advisor riding along. That’s more than double, at a sliver of the cost of running Opus outright.

Why this matters more for us than for coders

Here’s the thing I missed at first. Our work is more suited to this than coding is, not less. Anthropic’s own advisor docs name the sweet spot: research pipelines, data classification, document extraction. That’s our whole Content Factory.

Think about what a personal-brand audit actually is. Pulling the domain rating, the keywords, the reviews — that’s mechanical. Drafting the first pass of a section — that’s bulk. But the honest score, the “earned but illegible” insight, writing it in your voice — that’s the judgment, and it’s maybe five percent of the tokens. We were running all of it on the premium model. Like hiring a brain surgeon to take blood pressure.

The test is one question: does getting this wrong take taste, honesty, or somebody’s real voice? No — push it down to a cheap model. Yes — keep it on the ceiling, and let the cheap tiers tap it on the shoulder when they get stuck.

What we shipped

Three pieces, because a good idea that lives in one chat window dies in one chat window.

1. A standing rule the agent follows on its own

We wrote it into memory so the agent routes by tier automatically — no more waiting for someone to type “use Fable.” It picks the cheapest engine that clears the bar, delegates the bulk to subagents, and keeps the judgment on top. The full SOP is a document called Model Judgment & Delegation that anyone on the team can read.

2. A skill in every pack we give away

The Local Service Spotlight skill pack — the one on our spotlight sites and on dennisyu.com — now carries a model-judgment skill. It sits right next to boil-the-ocean, and together they say the whole thing: boil-the-ocean is finish the job, don’t stop at 90%; model-judgment is here’s how to power the job so finishing it stays cheap. Anyone who installs the pack gets both.

3. A real escalate-up path in the Audit Factory

Our batch audit engine already delegated down — scripts collect, a cheap model drafts, the frontier model judges. We added the mirror image: an opt-in --advisor flag that lets the cheap draft tier consult Opus the moment it’s stuck, instead of drafting blind and hoping the judge catches the miss. It’s off by default, so nothing about existing runs changed. It ships with tests.

The honest caveat

Cheap does not mean cheap on everything. The one place I’ve watched a model give itself away is voice. Ask a bargain model to write in someone’s real voice and it comes out smooth, safe, and obviously machine-made. So voice stays on Fable. The judgment score stays on the ceiling. And the honesty rules — the bands, the Knowledge Panel gate — live in code, not in a prompt, so no tier can hand you a flattering number. Cut corners on the grunt work all day. Never on the judgment.

Try it yourself — one sentence

You don’t need any of our machinery to start. Paste this into a Project, a memory file, or the top of a chat:

Then watch your premium budget stop evaporating by noon.

Read the full SOP
Get the skill pack

Sources: Simon Willison, “Fable’s judgement” (Jul 3, 2026) · Anthropic, the Advisor Strategy (Apr 9, 2026) · Anthropic, Create custom subagents. Part of the BlitzMetrics Content Factory. This article was itself built by a Claude agent that delegated its own bulk work down a tier and kept the writing on top — which is the whole point.