Draft a watch with AI

Writing a good condition and picking sources takes practice. The drafter does a first pass for you, and you edit it.

It never creates anything

The drafter returns a proposal. Nothing is saved until you review it and press save. That is the whole design: the model suggests, you decide.

Using it

1. Open the new-watch form in the admin panel or the Android app and choose “Draft with AI”.

2. Describe what you want, the way you would tell a person:

   Tell me when the 2026 examination result is declared. I have already applied, so I do not care about the application window.

3. Read what comes back. You get a suggested name, a condition, one to three sources with collection hints, and often a schedule and thresholds — plus a short rationale for each source and any assumptions it made.

4. Check the suggested domains. Domains it proposes to pin are liveness checked and flagged if they do not respond. A flagged domain is not removed — you decide whether it is a typo or simply a slow site.

5. Preview a search source. The preview runs the query once and shows the top results, so you can see whether it finds the right pages before saving.

6. Refine in one round if it is close but not right: “only official sources”, “check twice a day”, “ignore coaching sites”.

7. Edit and save. Everything is an ordinary form field — the draft has no special status once it is on screen.

What it is good at

• Turning a vague intent into a specific condition with exclusions.
• Suggesting a sensible pinned-official plus open-news source pair.
• Picking plausible starting thresholds and a schedule.

What to check every time

Check	Why
The pinned domains are real	A model can invent a plausible-looking domain. A wrong pin fails safe — it simply will not grade as official — but the source then finds nothing
The condition excludes what you do not want	Exclusions are where noise comes from
The schedule matches reality	Suggested schedules are guesses; and a watch cannot run more often than your instance does
The subject is precise	“the examination” may need to be “the 2026 cycle of examination X”

From the API

curl -s -X POST https://<your-host>/watches/draft \
  -H "authorization: Bearer $TOKEN" \
  -H 'content-type: application/json' \
  -d '{"description":"Tell me when the 2026 examination result is declared"}'

Refine by sending the previous draft back with an instruction. Full shapes: AI assistance API.

Limits

• 10 requests per minute per account across all AI routes.
• Returns 503 if the instance has no model key configured.
• Drafting is for new watches. To change an existing one, use Revise with AI — it proposes a field-level diff instead of replacing your hand-tuned sources.