Admin guide
Models & prompts
Each AI job runs on a model you pick and a prompt you can rewrite, both from Settings → AI. A model is a trade between accuracy and cost, and a prompt is the wording the job runs on, so between them the two controls decide how well a job does its work and what it costs to do it.
Where they areEvery job's model and prompt are on the AI settings page
Settings → AI lists every job the app can do. Each one shows what it does and what it sends, its own switch, a Model row, and a Prompt link. AI covers the switches. This page covers the two controls beside them.
Everything here is in the app. There is no file to edit and nothing to redeploy, and a change applies to the next run of that job.
The five jobsThree file things under a life area, two read and write a line
Three of the jobs file something you wrote under one of your life areas. Each takes a thing with no area, sends it along with your areas and what each one means, and gets back one area name or the word for none of them.
- Area naming
- Files a new journal entry under one of your life areas. Sends the entry's text and your areas with their descriptions.
- Task filing
- Files a task that has no life area, when no stream already answers for it. Sends the task's words, its kind and its note, with your areas.
- Document filing
- Files a document that has no life area. Sends the document's name, kind, description and text, with your areas.
The other two read something and write words back rather than picking from a list.
- Transcribe a recording
- Transcribes a video or a voice memo, then writes a title and a summary for it. The recording goes to ElevenLabs for the transcript, and that transcript goes to Anthropic.
- What changed in a version
- Writes the line under each version in History, saying what moved since the version before, and for a stream whether the target itself moved. Sends the two versions' text.
Choosing a modelEach job runs on the model you pick for it
The Model row offers Opus and Haiku, and the one in use is the outlined one. Click the other and it takes effect at once and stays picked. The choice belongs to that job alone, so you can leave the filing jobs on one model and move a different job to the other.
Opus is the model the app ships every job on. Whoever runs your LifeOps can point the whole instance at a different starting model, in which case a job you have never picked for runs on that one instead. Your own pick always wins over both.
Haiku is faster and cheaper and gets more of the judgements wrong, which matters most for the three filing jobs. A task filed under the wrong area ends up in the wrong pile rather than merely mislabelled, and nothing asks about it again.
Editing a promptPrompt opens the wording the job runs with
Prompt opens a box holding the instructions that job sends. It arrives with the wording the app ships. Change it, press Save, and that job runs on your wording from then on. The link then reads Prompt — edited, so you can see at a glance which jobs you have changed. Close prompt shuts the box without saving.
Reset to default appears beside Save once a job has an edit, and puts the shipped wording back.
Two things steer a filing job, and the prompt is the second of them. Each of your life areas can say what it means, in a line under its name, and the three filing jobs are sent those lines every time they run. Writing that line is the smaller change and it applies to one area. Editing the prompt changes how the job reads every area at once. Settings in the user's guide covers the area descriptions.
What gets refusedAn edit that would break the job is not saved
The three filing prompts hold markers in curly brackets where the app drops the job's inputs in — the list of your areas, and the thing being filed. Save a prompt with one of those markers missing and the app refuses it, naming the ones it needs, and your text stays in the box for you to fix. While you are editing one of those three, a line under the box names the markers that have to stay in.
A stray curly bracket is refused the same way. Write a literal brace as a doubled one.
The two reading jobs take their material as a separate message rather than inside the prompt, so their wording has no markers it must keep.
With AI offA switch keeps your picks, and a missing key takes the page away
A job you switched off keeps its model and its prompt, and picks them back up when you switch it on again. The master switch above the list works the same way, with every job's controls still on screen and none of them running.
An instance whose Anthropic key has been taken off the backend is the other case. The AI page then replaces the whole list with one sentence saying AI is switched off for the instance, so there is no model to pick and no prompt to edit until the key is back. AI covers what a job's own switch stops, what the account's switch stops, and what a missing key stops.