processes & runsmini guide · 6 min

make a process

turn something that worked once into a recipe a pixie runs again and again, with the inputs you choose.

A process is a canvas you can run again. You make something once, by hand or with a pixie. Then you name the parts that change each time, such as a subject, a reference picture or a style. From then on, every run is that same canvas worked again by the same pixie, with new values.

In this guide you turn a working canvas into a process, then start its first run.

before you start

  • A canvas where the work already works. The process copies what is on it, so get one good result first. Here that's a monogram: an inspiration picture goes in and a clean SVG monogram comes out.
  • A pixie on that canvas that knows how the work was done, ideally the one that did it.
  • The pixie set to Coaching or Executing. In Brief mode a pixie researches and plans only, so it can't build a process. Switch modes from the pixie's chat.

make it

  1. Ask the pixie to make it a process. In the pixie's chat, say what should change from run to run: "make this a process — the inputs are an inspiration image and the letter." The pixie packages the canvas as a process, names it, and declares those inputs.

  2. Meet the resident. A process node appears on your canvas, and the pixie moves in as its resident. The first process a pixie claims gives it a craft name, so Suki becomes something like Suki the Monogrammer. From now on that pixie works every run of this process.

  3. Check the inputs. The node's New run form has one field per input, each with its own handle on the left. Inputs can be text, an image, a number, a yes/no, or a pick-one or pick-several list. A process can have up to 8 inputs. To change a label or add a choice, ask the resident; it can revise the inputs at any time.

  4. Fill the form and press Run. Run starts a run with exactly these values. The run gets its own canvas, cloned from the process, and the resident starts working there. You can close the tab; the run keeps going.

  5. Watch it, or don't. The run appears in the node's Runs table, with a spinner while the pixie works. Click its name to enter the run's workspace and watch, or chat with the pixie there. When the pixie has something for you, the node shows a dot.

  6. Review what comes back. By default, results wait for you. With human review set, they are staged for your approval and publish the moment you approve. Approved results appear in the run's Results cell and in the node's tray, ready to wire onward.

choose how results are reviewed

The chip in the Runs header sets how this process's runs deliver:

optionwhat happens
auto approveresults publish the moment the pixie submits them, with no review
human reviewresults are staged until you approve them; approving publishes them instantly (the default)
self reviewthe pixie critiques its own results and retries before publishing on its own
self + humanthe pixie self-reviews first, then results wait for your approval

Until you set it, a process uses the canvas default. More on review: review and improve.

run it again

  • Several at once: Add row puts a draft run in the table. Fill in a few rows, then press each one's Run. A draft costs nothing until you run it.
  • Same inputs, new run: Run again on a row starts a new run with that row's values.
  • See the recipe: Template opens the process itself, the canvas every run starts from. Changes you make there apply to the runs you start afterwards.

what just happened

The process is a template canvas plus a resident pixie. Each run is a full copy of the template, so a run can't spoil the recipe, and one run going sideways never touches another. Because the resident reads the inputs as a brief rather than plugging them into fixed slots, a process can handle work that needs judgment: picking the better of two drafts, asking you a question halfway, or delivering three results instead of one.

take it further

  • runs — the runs table, run workspaces and what "finished" means.
  • review and improve — reject with notes, and fold a good run back into the template.
  • build a pipeline — chain processes so one's results feed the next.
  • derived columns — run a process once per row of a table.

checked against the product · 2026-10-23