Lean body that assumes Claude's competence — it opens with the unifying mental model ('Every cost question reduces to an aggregation...') and adds only PostHog-specific facts (event names, double-count pitfalls) Claude wouldn't already know, with no boilerplate concept explanations.

Provides fully executable, copy-paste-ready artifacts: a HogQL query for total spend, a JSON tool-call payload for a single trace, a tool-purpose table, and exact UI link formats; ready-to-run recipes are clearly signaled in the breakdown-patterns reference.

Multi-step workflows are clearly sequenced (total spend, breakdowns, single-trace, regression, materialize) with explicit validation guidance — 'Always set a time range', 'read $ai_cost_model_source first when a cost looks wrong', and a 5-step regression playbook referenced for fragile debugging.

The body is an overview that signals six one-level-deep references via clear inline links plus a References section, and all referenced paths (cost-properties, cost-sources, cache-accounting, breakdown-patterns, regression-debugging, materializing) resolve to real files in ./references/ with no nesting.

Lists multiple concrete actions — 'total cost over time, cost by model, provider, user, trace, or custom dimension, token and cache-hit economics, and cost regressions' — each tied to a concrete aggregation rather than vague language.

Explicitly answers both what (the cost investigations listed) and when via a clear 'Use when the user asks ... or pastes a trace URL and asks about its cost' trigger clause.

Quotes natural user phrasings verbatim — 'how much are we spending on LLMs?', 'which model / user / feature is most expensive?', 'why did cost spike?' plus trace-URL and dashboard/alert mentions — covering the terms users would actually say.

Scoped to PostHog LLM-spend analysis over $ai_generation/$ai_embedding events with distinct triggers (spend questions, cost dashboards, trace URLs), giving it a clear niche unlikely to conflict with sibling skills.