Goals

The goal is what the AI change judge uses to decide whether a page is changed vs same. Convert the user’s intent into a concise 2-3 sentence goal:

• Start with Alert when ... and state the trigger using the user’s wording.
• Restate any scope they mentioned: top N, price, role type, region, company, topic, status, or a specific entity.
• Add an Ignore ... sentence only for intent-specific exclusions (e.g. points/comments for rankings, marketing copy for pricing, general company-page updates for job listings). The judge already handles generic noise — whitespace, casing, punctuation, encoding, formatting-only changes, request/session IDs, cache busters, tracking params, generic metadata, and unrelated page chrome — so leave those out.
• Include only page-specific sections, entities, thresholds, exclusions, or business rules the user actually mentioned.
• If the user is vague or asks for “any change”, keep the goal broad with no exclusions. If the user mentions noise they do not care about, include that explicitly.

Writing good --queries (web monitors)

For a web monitor, queries control recall (what the search retrieves) and the goal controls precision (which results alert). Tune both — a perfect goal can’t alert on a result the queries never pulled in, and broad queries with a vague goal produce constant low-value alerts.

• Write keywords, not sentences: OpenAI new model release, not tell me when OpenAI releases a new model.
• Quote multi-word entities ("Llama 4"); group synonyms with OR (launch OR release OR announcement).
• Keep each query tight (~2–6 terms). One broad query usually beats several narrow ones — extra queries split the --max-results budget without adding coverage.
• One query per distinct subject. Several facets of one subject = one query; only split for genuinely separate entities (e.g. “OpenAI, Anthropic, and Google”).
• Restrict or exclude sources with --include-domains / --exclude-domains rather than site: operators in queries.
• --search-window sets recency — 5m, 15m, 1h, 6h, 24h, 7d (default 24h). Widen it for niche topics that don’t publish often.
• --max-results caps results per query, 1–50 (default 10).

What good looks like: a healthy web monitor mostly returns new: 0 and alerts only on genuinely new, on-goal results. If many retrieved results are off-goal, the queries pull noise the goal rejects — tighten the queries. If a topic returns nothing for long stretches, the queries are too narrow or --search-window too tight — broaden them. If the user dismisses alerts, the goal is too broad — add an intent-specific Ignore .... The aim is high precision with enough recall: every alert worth acting on, nothing real missed.