Two tools with clear naming and documented input schemas, but output schemas are undocumented and error handling lacks recovery guidance. search_conferences has good parameter descriptions (min_date, max_date, country, tags, cfp_open all clearly annotated). apply_conferences is less complete, it lacks explicit output schema documentation and the error handling returns a generic dict without recovery hints. Both tools follow verb_noun naming (search_, apply_) which is good. The main gap: no documented return schemas, no error classification or retry guidance, and apply_conferences has incomplete parameter descriptions for returned conference_summary structure.

• Document the output schema for search_conferences: specify that it returns a list of conference dicts with fields {name, tags[], location, cfp_deadline, city, country}. State the maximum result count (e.g., capped at 50 results) to prevent context window exhaustion.
• Document the output schema for apply_conferences: specify the structure of the success response (e.g., {matches: [{conference, match_score, reason}], applied_count, errors?}). This enables the LLM to extract and act on the data.
• Enhance apply_conferences description to explain when to use it vs search_conferences: 'Use this to intelligently match a talk to conferences based on content similarity. Requires a talk resource URI (file:// or talk:// scheme). Returns ranked matches with explanations.' This disambiguates from search.
• Add error recovery guidance to apply_conferences. When the talk_resource_uri fails, return: 'Unable to read talk from {uri}. Available talks: [list_talks]. Use talk://theme_name or file://absolute_path.' This guides the LLM to self-correct.
• Add pagination/limit state to search_conferences description: 'Returns up to 50 matching conferences. For larger result sets, use filters (date range, country, tags, cfp_open) to narrow results.' This prevents context bloat.
• Add parameter relationship documentation: In apply_conferences, clarify that cfp_open is always set to True internally (conferences must have open CFPs), so the user cannot override it. Or expose it as a parameter with explanation.
• Categorize errors: 'talk_resource_uri not found' is user-fixable (call search_resource or list_available_talks); network errors reading the resource are retryable. State this in error responses so LLM knows what to do.