Three tools with basic schemas and minimal descriptions. All tools have input schemas with proper JSON types, but descriptions are generic (10-60 chars) and lack context for LLM selection. No parameter descriptions beyond names. No output schema documentation. No error handling guidance. Missing idempotency hints for record_match (a write operation). Tool composition is reasonable (three distinct concerns), but descriptions do not explain WHEN to use each tool or what the LLM should expect in responses.

• Add parameter descriptions to all tools. For record_match: 'candidate_a (string): Unique identifier for the first candidate (e.g., prompt name, agent ID). Must match identifiers used in previous matches.' For score_a: 'Score for candidate A (float, 0.0 - 1.0): 1.0 if A wins, 0.5 for draw, 0.0 if B wins.'
• Document output schemas in tool definitions. E.g., record_match returns {candidate_a, candidate_b, score_a, old_rating_a, old_rating_b, new_rating_a, new_rating_b, delta_a, delta_b}. get_leaderboard returns array of {rank, candidate, rating, matches}.
• Expand tool descriptions to 80 - 150 chars with WHEN/WHY context. E.g., 'Record a match result to update Elo ratings. Call this after comparing two candidates (e.g., LLM outputs, prompt strategies). Ratings adjust based on expected vs. actual outcome.'
• Add error handling with actionable messages. Validate score_a in [0.0, 1.0] and return: 'Invalid score_a: {value}. Must be 0.0 (B wins), 0.5 (draw), or 1.0 (A wins).' Validate candidate identifiers and suggest available candidates if not found.
• Declare record_match as destructive with idempotency guidance. Add tool annotation: destructiveHint=true. Document: 'This operation modifies ratings permanently. Retrying with identical inputs will apply the match twice, use a match_id parameter or external deduplication to ensure idempotency.'