Mods

route

Feature and test work goes to the cheaper model by itself.

It comes with the pack and stays off until you turn it on.

The moment

You type: implement pagination for the users endpoint.

Without route

Claude reads the code and builds the feature itself, step by step. A hand-off would have been cheaper, but only you can ask for one, and you would have to remember every time.

With route

route asks a small, fast classifier what kind of work each prompt is. Only feature or test work gets the order, in your name, to hand it to delegate’s worker before reading the code. A question, a rename or a typo fix stays with Claude, as before.

Pick a prompt and see what route does with it. The labels and scores are the ones the classifier gave in our test.

Pick a prompt

You typeimplement pagination for the /users endpoint

The classifier saysfeaturescore 1.00, needs 0.7

The order goes beside your prompt: hand it to delegate’s worker before reading the code.

The order, word for word
This prompt is feature work, and the person's setting for feature work is: implement it through the hand_off tool (kind build). Hand it off before you read the code: the worker reads it, so write the task from the request as given. Write the request into the task so it stands alone, with the project's whole test suite as the check. The tool checks the worker's work itself; when its result says Verified and the diff is what was asked, report to the person without running anything again.

What we measured

On click’s eight jobs it made the whole set 14% cheaper14.2% cheaper, 95% range 10.1 to 17.9% cheaper14.2% cheaper
95% range 10.1 to 17.9% cheaper
32 runs, 4 per task
Opus 5.5, classifier ScaleDown, 15 Oct 2026 than the pack alone. On a TypeScript feature, 51% cheaper51% cheaper51% cheaper
No range in the source
4 per task and condition
Opus 5.5, worker Sonnet 5.5, 15 Oct 2026, with Claude taking two steps of its own instead of eleven11.0 to 2.011.0 to 2.0
4 per task and condition
Opus 5.5, worker Sonnet 5.5, 15 Oct 2026. Both classifiers labelled all 5050 of 5050 of 50
50 calls each
TypeSafe Jev and ScaleDown Classify, 15 Oct 2026 of their test calls right.

Every measurement for route

The dot is the measured change, the bar around it the 95% range, the upright line no change. Left of the line is cheaper or fewer.

Where it doesn’t help

It is off until you turn it on, because the start of each prompt goes to an outside classifier in the US. Prompts under 40 characters are never sent, so a short “write tests for the parser” stays with Claude. And every prompt in our tests was a complete specification; on a vague request, a task written before reading the code may be a worse task. We haven’t tested that.

Turn it on or off

Turn it on with /mods route on and a key for TypeSafe’s Jev or ScaleDown. It needs delegate. Every option is in the README; the code is in route/.