The build brief — what you are trying to build

Or pick a file: it is read locally, nothing uploads until you run.

Constraints — the numbers and limits, if you have them

Where this is

What you want out of it

Prompt-cache arithmetic — free, and the one number here that is not an opinion

A cache write costs more than a plain pass and a cache read costs a fraction of one, so caching a shared prefix only pays once enough requests reuse it inside the window. This works the ratios for your numbers, in your browser.

Requests sharing the prefix, per cache window

Cache lifetime

Share of the request that is the stable prefix (%)

The prefix is whatever is byte-identical on every call: the system prompt, the tool list, a fixed corpus. Anything that changes per request is the suffix and is paid in full either way.

The route

What this brief actually is

The recipe stack

#	Recipe	Role	Why

Recipe by recipe

Core first, in build order. Each card carries the phrase from your own brief that put it here — if that quote is not in your brief, the routing is guessing and you should say so.

Ruled out — what looks right for this brief and is not

Prescan coverage

Eval plan How you will know it works. A route with no way to tell is half a route.

#	Check	Passes when

Where the cost goes

Open questions — answer these and re-route

Build sequence

Summary

Result

How it works

Nothing to hand? Load the — a high-volume triage problem where the obvious answer is retrieval and the right answer is mostly not — or the , where the documents are PDFs, the output feeds a downstream system, and the interesting question is what you can prove about the extraction.

Write the brief — the prescan is free

No upload, no AI: the prescan reads your brief in this browser and looks for signals of all twenty-one cookbook recipes — classification, retrieval, summarization, tool use, SQL, vector stores, embeddings, vision, charts, forms, PDFs, sub-agents, evals, JSON mode, moderation, prompt caching and the rest — showing you the exact sentence that triggered each one. It also names what your brief does not say: no volume, no latency budget, no accuracy bar, no cost ceiling, nowhere the data lives. Those gaps become questions the paid pass has to hand back to you. The cache calculator in the same panel is real arithmetic, not a model opinion.

The router picks the stack — this is the metered part

A staff engineer’s pass over your brief: the shape first (one call, a workflow, an agent, a batch job, or a cheap-lane-plus-escalation hybrid), then an ordered stack of recipes with a role of core, supporting or optional. Every entry has to quote a phrase from your own brief to justify itself — no quote, no recipe. Just as important is the rule-out list: the recipes that pattern-match to your words and are the wrong call, with the reason. Pricing is honest — a worst-case amount is reserved before the run and only what the run uses is charged; the meter beside the button shows both.

Answer the open questions and route again

The route comes back with an eval plan that has a pass threshold per check, cost notes that name which lane in your stack dominates spend, a build sequence, and the questions your brief left open. Answer two of them in the constraints box and route again: the second pass is a different plan, not a reworded one, because the volume and the latency budget are usually what decides between a workflow and an agent. Everything exports — the stack as CSV, the whole plan as Markdown or JSON, and a paragraph you can paste straight into a design doc. Routings are saved to your SkillSafe account when you are signed in, and restoring one puts the brief back in the form.

Derived from the @anthropics/claude-cookbooks skill — the recipe catalogue this routes across.