The body is lean and assumes Claude's competence with no padded concept explanations, but the 'ask the user / present the list' guidance is restated almost verbatim across step 3, step 4, and the Error Handling table, so some tokens do not earn their place.

Provides concrete, executable guidance — a capability-to-question table, explicit parameters to pass (file path, line number, source scope, observation window), and specific selection heuristics; the absence of fixed code is justified by runtime tool discovery.

A clearly sequenced six-step flow with explicit checkpoints ('ask the user before proceeding — do not guess', preflight pass/fail criteria) and an error-recovery table; the read-only operations do not trigger the destructive/batch validation cap.

The body is an overview that points to one real, one-level-deep reference (references/mcp-tool-discovery.md, with a signaled anchor) holding shared discovery rules, keeping the skill-specific flow inline and the shared detail external.

Lists multiple specific concrete capabilities — 'current variable values, execution durations, hit counts, and value distributions' and 'instrumenting running services with Lightrun MCP tools' — rather than vague domain language.

Explicitly answers both what ('Answer questions about live production system behavior…') and when ('Use when the question requires live runtime data rather than static code analysis'), with concrete trigger examples.

The 'Use when' examples ('show recent requests to this endpoint', 'which branch runs for customer X?', 'what values appear for this expression in production?') are natural phrases a user would actually say when they need this skill.

Clear niche — live production runtime instrumentation via Lightrun — explicitly contrasted with 'static code analysis', making it unlikely to trigger for the wrong skill.