working modes

A pixie's working mode decides what it is allowed to do without asking you. There are three you can pick, Brief, Coaching and Executing, and a fourth, Review, that a run pixie enters on its own after you reject its work.

Switch modes with the pills in the header of the full chat thread (double-click the pixie, or use the expand icon in the chat). The change applies to your very next message.

the three modes

mode	what the pixie does	who presses Run
Brief	researches and clarifies only; agrees on a plan	nobody yet: it has no build tools
Coaching	builds, wires and configures; you steer it	you
Executing	runs models itself and iterates before showing you	the pixie

A new pixie on a shared canvas starts in Brief.

brief

The pixie asks question cards, researches models, checks your org wiki, looks at what is already on the canvas and proposes an approach as a plan card with the models, wiring, key settings and what success looks like. It confirms with an approval card.

In Brief the pixie can't add, wire, configure or run anything. Once you approve the plan, the chat switches itself to Coaching.

coaching

You are teaching the pixie the process. It asks about what is unclear, then builds: it adds nodes, wires them and sets the inputs. It can't run a model itself. When a node is ready it pins a yellow RUN tab on it and tells you what pressing Let's go! will produce.

Coaching also watches what you change. If you edit a prompt or swap an input in the pixie's working set, it takes your change into account rather than overwriting it. It keeps asking whether each experiment made things better, and records what worked in its memory.

executing

The pixie runs model nodes itself. Several independent generations go out in one batch. After each run it looks at the result and judges it against your request.

• Up to 3 attempts per deliverable. If the third still falls short, it shows you its best and says plainly what fell short.
• If attempts are drifting further from the goal, it stops early and talks to you.
• If something genuinely blocks it, it stops and asks instead of finishing the whole job on a guess.

review

Review is not a mode you pick. A pixie that works the runs of a process enters it when you reject the work it submitted. The mode pill shows Review and is locked.

In Review the pixie:

1. reads your feedback word for word, and asks what fell short if you declined without a reason;
2. says which stage produced the miss and why;
3. reworks that stage and resubmits;
4. once you approve the fix, asks whether the process itself is at fault, and with your click files a report to the process's owner.

See review and improve.

which one to use

• A new idea or an unfamiliar request: Brief. Get the plan right before anything is built.
• A task you want to shape as you go, or a process you are training: Coaching.
• A well-understood job, or many variations: Executing.
• Cost-sensitive work: Coaching. Every run is yours to press, so nothing is spent without you.

limits & good to know

• Residents are always in Executing. Run pixies, sandbox devs, modelers and designers show a single locked pill; posties and keepers show no mode pill at all.
• Executing spends credits as the pixie runs models; in Coaching every model run is yours to press.

what a pixie turn costs

ixi picks the language model each pixie thinks with; you don't choose it.

A turn is everything the pixie does between your message and its answer: reading, several tool calls and the reply. It is billed once, when the turn ends, for the tokens used plus a margin of 15% by default. The charge comes out of your organisation's credits, next to the models the pixie runs for you.

• The chat header shows a running token count, ↑ for input and ↓ for output. Hover it for the exact figures.
• Pixie turns show in your organisation's Usage tab with a chat icon beside the model; the balance is in Billing. See billing and credits.
• When your organisation is out of credits the pixie answers with a normal message saying so, with a link to the billing panel.
• Pixie turns don't count toward the 500-runs-a-day limit on model runs. See model nodes.

related

• chat
• skills and tools