The composer

The composer is the field at the foot of the roster column. Write a sentence describing what you are building, and it returns the skills for it — as a proposal you read first, not a selection it makes for you.

It is the route for a project that does not exist yet. There is no code to detect anything from, and you should not have to learn a catalogue to get started in it. For a project that does exist, the better route reads the code: Adding to an existing project.

Quick start

1. Sign in with GitHub. The composer calls a model, so the cost has to be attributable to somebody; the rest of the editor works signed out.
2. Type a sentence into the field at the foot of the roster — “a Next.js app with Postgres and Playwright tests” will do.
3. Press Send, or ⌘↩ (Ctrl↩ off a Mac). A plain Enter inserts a newline — the field takes prose.
4. Read the proposal that appears above the field, then Apply or Discard.

Nothing changes until you apply. Sending is not applying, and neither is discarding.

What comes back

A list of skills, and one short sentence saying why. That is the whole answer.

Each row names a skill and the load it will rest at once selected. Skills you have already selected are dropped from the proposal rather than listed — proposing something you already have reads as a change that will not happen.

Apply selects them. Discard removes the block and leaves your sentence in the field, because the usual next move is rephrasing rather than starting again.

What the model is not allowed to decide

Almost everything, and this is the design rather than a limitation.

The model returns skill ids and a sentence. It says nothing about scope, install mode, which sub-agents carry what, whether each copy is preloaded, or any sub-agent’s model and reasoning effort. Those are derived from the same rules the CLI generates from — so a model with an opinion about them would produce a configuration the CLI then contradicts. Two sources for one answer is how they come to disagree.

Its ids are not trusted either. They are filtered against the catalogue before the editor draws anything, so a plausible near-miss — an id that looks exactly like one of ours and is not — is silently absent rather than a broken row.

Applying goes through the same verb a click does. Each proposed skill is selected the way clicking its cell selects it, one at a time, so the incompatibility rules, the implied skills and the exclusive-category swaps all run exactly as they would for a click. A proposal cannot reach a configuration you could not have reached by hand — it is a faster route to the same grid, not a second way of writing to it.

Editing the sentence voids the answer

A proposal belongs to the sentence it was asked for. Change the sentence and the proposal goes, because a stale answer under a changed question is worse than no answer.

That holds mid-flight too: if you rephrase while a request is still running, the answer is dropped when it arrives rather than drawn under the new sentence.

Nothing here is persisted. The draft and the proposal are local to the page — they do not survive a reload, and they are not part of your configuration until you apply them.

When it refuses

Five endings. Signing in and shortening the sentence are the two you can act on.

What it says	What happened
Sign in to use the composer — nothing was sent.	You are signed out. Nothing left the browser.
Too many requests in a minute. Try again shortly.	Ten calls a minute, per person. Wait.
That is too long to send. Shorten it and try again.	Over 600 characters. Refused before the model was called.
The model did not answer. Nothing changed.	The call failed upstream.
Could not reach the composer. Nothing changed.	The network did.

The 600-character cap is a real limit rather than a nicety — input tokens are billed, so an unbounded prompt is an unbounded bill.

One refusal never leaves the browser at all: a blank sentence. Send is out of reach while the field is empty or holds only spaces, and ⌘↩ refuses on the same check.

What it is not

It is not a chat. There is no thread, no history and no follow-up turn. One sentence, one proposal, and the proposal is a changeset rather than a reply.

It is not the only way to change a selection. Anything the composer can do you can do in the grid, and several things only the grid can do — routing a skill to one sub-agent, pinning a model, adding a skill from outside the catalogue. Selecting skills covers those.

It does not install. Like everything else in the editor it ends at the install dialog and a command you run — Installing and sharing.

Where to go next

• Four ways in — the composer beside the other three routes.
• Selecting skills — the grid the composer writes into.
• Adding to an existing project — the same job when there is already code to read.