collibra-atlas: Skill for Claude Code

Web Browsing

The web is a firehose. The skill is not in finding information — it is in finding the right information and knowing when you have enough.

Tenets

• Search is iterative. The first query rarely returns the best result. Every page you read teaches you better terms for the next query.
• One query, one intent. A query that asks two questions answers neither well. Split "Python async vs sync performance and Django ORM optimisation" into two searches.
• Source authority matters. A blog post from 2019 is not evidence for a 2026 claim. A corporate marketing page is not an independent review. Factor the source into every fact you extract.
• Corroborate before citing. A single source is a lead, not a fact. Two independent sources agreeing is a fact. Three is confirmation.

Leading words

• Scout — A broad, exploratory search to map the territory. Before you commit to deep reading, know what exists. Scout searches use wide queries and skim results.
• Sift — Evaluate and filter what you found. A result's rank does not equal its value. Sift by relevance, authority, freshness, and independence before deciding what to fetch.
• Corroborate — Cross-reference each claim against a second independent source. A claim that cannot be corroborated is a claim you qualify ("According to one source…").

Phases

PHASE 1: Plan

Completion criterion: the question(s) to search for are written down, each as a separate query string, and the appropriate search backend is chosen.

1. Decompose the user's request into independent questions. Each question gets its own search.
2. For each question, choose the backend by the nature of the answer needed (see BACKENDS.md).
3. If the question refers to a specific known URL (docs page, blog post, GitHub repo), skip directly to Phase 4 (Extract) — no search needed.

PHASE 2: Scout

Completion criterion: search results returned for every question. At least one result per question looks promising enough to fetch.