Skip to content
Watcher Docs

Conditions

A condition is a sentence describing what would count as news. Watcher shows a language model your freshly collected text and this sentence, and asks: is this condition met?

condition: >
  Alert when a new stable version of the framework is released. Ignore
  release candidates, beta announcements, and blog posts about future plans.

The rules the model works under

Section titled “The rules the model works under”

You are not writing a prompt into a vacuum. Watcher wraps your condition in instructions that constrain the answer:

  • Only clear, confirmed evidence counts. Speculation, expectations, predictions, rumours and “coming soon” must not trigger.
  • A verbatim quote is mandatory. If the model triggers, it must copy the exact sentence that proves it. If it cannot find one, it is told not to trigger.
  • Restatements do not count. Watcher passes a short history of what it has already told you; an item repeating a known event is not a new event.
  • Page text is data, never instructions. The model is explicitly told to ignore any text inside a collected page that tries to give it new rules or claim something is “confirmed”. A page cannot talk Watcher into alerting.

Those constraints are why a vague condition tends to under-fire rather than spam you.

Write it like a brief, not a keyword

Section titled “Write it like a brief, not a keyword”
Instead of Write
result Alert when the official result or scorecard for the 2026 examination is declared. Do NOT alert on exam dates, admit cards, answer keys, or coaching articles.
price drop Alert when the listed price of this specific product is lower than £300. Ignore unrelated products, coupons, and "sale starts soon" banners.
security Alert on a newly published security release or CVE affecting this project. Ignore general security commentary and unrelated advisories.

State where you already are

Section titled “State where you already are”

If you have already done part of a process, say so. It stops Watcher alerting on stages that are behind you:

condition: >
  I have ALREADY applied. Trigger only on a genuine new post-application
  milestone: a correction window, the exam date, the admit card, the answer key,
  or the result. Do NOT trigger on the application window opening or closing.

Length and specificity

Section titled “Length and specificity”

A condition of two or three sentences is usually right. Longer conditions are fine — there is no practical limit for normal use — but each extra clause is another thing the model can misread. Be specific about the subject (which product, which exam, which project) so a similarly-worded story about something else cannot match.

What happens to the answer

Section titled “What happens to the answer”

The model returns a verdict: triggered or not, a confidence between 0 and 1, a one-line summary, the quote, and often the date of the evidence. That verdict is then checked — the quote must really appear in the collected text — and only then does the tiering logic decide whether you hear about it.

Going deeper

Section titled “Going deeper”