LLM Cost & Latency Budget Skill

LLM features have a unit cost and a tail latency that demos hide and production exposes. This skill does the token math up front — what one request costs, what a million cost, where the p95 latency comes from — and lays out the levers (model tiering, caching, prompt trimming) so cost and speed are designed, not discovered.

Required Inputs

Ask for these only if they aren't already provided:

• The request shape — typical system prompt, user input, retrieved context, and output sizes (in rough tokens).
• Volume — requests/day now and at target scale; peak concurrency.
• Models in play — candidate model(s) and their per-token input/output prices.
• Targets — acceptable cost per request (or per user/month) and the latency users will tolerate (p50 / p95).

Output Format

Cost & Latency Budget: [feature]

1. Per-request token math — a table estimating tokens in/out per call, and the resulting cost at each candidate model's price.

2. Monthly projection — per-request cost × volume, at current and target scale; the headline number leadership will ask for.

3. Model tiering — route easy requests to a cheaper/faster model and only escalate hard ones (cascade); show the blended cost. Often the single biggest saving.

4. Latency — where the p95 comes from (model TTFT + output length + retrieval + network), the target, and how streaming changes perceived latency even when total time is unchanged.

5. Cost levers — ranked by impact: prompt/context trimming, caching (prompt cache + response cache for repeats), shorter outputs (max_tokens), batching, tiering, and "do you need the model at all for this path."

6. Guardrails — per-user / per-day rate limits, a max-tokens cap, a spend alert threshold, and a kill switch — so a bug or abuse can't produce a surprise invoice.

Quality Checks

• Token estimates are itemised (system + context + input + output), not a single guessed number
• The monthly cost is projected at target scale, not just today's volume
• Model tiering / cascade is considered before accepting the flagship-model cost everywhere
• p95 (not just average) latency is targeted, and streaming is considered for perceived speed
• Caching is evaluated for repeated prompts/contexts
• A spend alert + rate limit + kill switch are specified to cap the downside

Anti-Patterns

• Do not budget on average latency — users feel the p95, and the tail is where AI features feel broken
• Do not default every call to the most capable model — most requests don't need it; tiering often cuts cost by more than half
• Do not forget output tokens cost more than input — verbose responses are often the hidden cost driver
• Do not ship without a spend cap and alert — an unbounded LLM feature is an unbounded bill
• Do not optimise cost before measuring it — itemise the real token usage first, then pull the biggest lever

Based On

LLM production cost/latency practice — token accounting, model cascades/tiering, prompt & response caching, and tail-latency budgeting.