Skip to content

Revise a watch with AI

Drafting creates a new watch from scratch. Revising changes one you already tuned — without throwing away everything you tuned.

Draft Revise
Works on A new watch An existing watch
Returns A whole draft Only the fields that should change
Risk to your work None (nothing exists yet) None — unmentioned fields are preserved by construction, not by careful prompting
  1. Open the watch in the panel or the app and choose “Revise with AI”.

  2. Say what should change, in one sentence:

    Check twice a day instead of hourly, and stop sending unconfirmed alerts.

  3. Read the diff. Each proposed change is shown as before → after, with the model’s one-line reason. Every row has a switch.

  4. Turn off anything you disagree with. Only the rows you leave on are applied.

  5. Save. The result is validated as a complete watch before it is stored, so an invalid revision can never be saved.

name, condition, schedule, threshold, weakThreshold, minSources, and sources.

Row shows Meaning
A value on both sides Change this field
An after value of “removed” Remove this optional field, e.g. dropping the weak tier
A field you did not expect Turn it off — partial approval is fully supported

A revision that proposes nothing means the model judged the watch already satisfies your instruction. The notes field usually explains why.

  • “Only alert on the official site, not news coverage.”
  • “Also watch for the answer key, not just the result.”
  • “Be stricter — I am getting too many false alarms.”
  • “Check every six hours instead.”
  • “Stop sending unconfirmed alerts.”
Terminal window
curl -s -X POST https://<your-host>/watches/<id>/revise \
-H "authorization: Bearer $TOKEN" \
-H 'content-type: application/json' \
-d '{"instruction":"check twice a day and stop sending unconfirmed alerts"}'

Nothing is written. Apply what you approve with a normal PUT /watches/:id. Shapes and the base parameter (for revising unsaved form state) are in AI assistance API.