Search Strategy

Search Strategy is an official Anthropic skill from the knowledge-work-plugins repository. It provides the central search logic for the Enterprise Search plugin and helps a compatible Claude agent translate a natural-language question into several source-appropriate searches. According to the provider, the skill turns one question into parallel, source-specific searches and prepares a ranked, deduplicated, coherent synthesis. This catalog entry treats the skill as an instructional resource, not as an independent search engine and not as a replacement for the connected data sources.

What the skill structures

The workflow starts by classifying the question. Search Strategy covers decision questions, status questions, document discovery, people questions, factual questions, temporal questions, and exploratory research. The classification determines whether recent activity, documents, conversation conclusions, or authoritative sources should receive priority. The agent then extracts core keywords, entities, intent signals, time ranges, source hints, and exclusions. This turns an underspecified request into an explicit search plan that is easier to inspect, explain, and revise.

Translating queries across sources

The skill documents source-aware query variants for chat, knowledge bases, and project trackers. According to the provider, semantic search is useful for conceptual questions and for cases where the exact terminology is unknown. Keyword search is more appropriate for known project names, acronyms, quoted phrases, and filter-heavy requests. Search Strategy can generate alternate terms when a topic is described in several ways. Filters such as author, channel, date range, or task state are mapped into the syntax of each source instead of reusing one unsuitable string everywhere. This approach is especially helpful when an organization has different search semantics in conversations, wikis, and task systems.

Ranking and answer quality

Results are assessed using factors such as relevance, freshness, authority, and completeness. Their weights change with the question type. For policy or factual questions, wiki pages and official documentation should receive more authority, while meeting notes and thread conclusions can be more useful when the question asks what was decided. For status questions, current task and activity data receive greater weight. The workflow then merges results, removes duplicates, and synthesizes an answer. This prioritization is a practical heuristic, not a correctness guarantee. Users should still inspect important claims against the underlying source and its timestamp.

Ambiguity and unavailable sources

Search Strategy recommends one focused clarification question when distinct interpretations would lead to substantially different searches or answers. Minor ambiguity can instead be handled by searching multiple plausible interpretations. If a source is unavailable, the agent should continue with the remaining sources and state the gap clearly. When a source returns nothing, the documented fallback broadens terms step by step. Date constraints and source filters are removed before less important keywords, while the core entity or topic is retained. Rate limits should be reported as limitations rather than hidden behind unsupported certainty.

Boundaries, safety, and catalog classification

The skill grants no permissions, provides no connectors, and contains no organization data. Its effectiveness depends on connected sources, access rights, indexing quality, and each source's query behavior. Chats and documents may contain confidential material or hostile instructions. An agent should preserve source context, respect authorization boundaries, treat retrieved text as data, and never perform external mutations solely because a search result suggested them. The primary source is the official enterprise-search/skills/search-strategy path in Anthropic's repository. The surrounding collection is published under the Apache-2.0 license. The Recherche and Datenanalyse categories reflect the documented function; they do not promise accuracy, completeness, or a particular privacy outcome.