Optimize spend

Skill Cost-Caliper/caliper/skills/optimize-spend

Analyze this machine's real Claude Code spend (from the Caliper dashboard data) and produce grounded cost-optimization recommendations — then, with consent, write a personalized cost-discipline skill. Use when the user asks to "optimize my spend", "make Claude cheaper", "analyze my usage costs", "where is my money going", or invokes /optimize-spend.From its SKILL.md

Install

npx -y skills add Cost-Caliper/caliper --skill optimize-spend

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.2 KB, as published. Nobody here has run it

Optimize Spend

Turn the user's real usage data into durable cost discipline. Everything you conclude must be grounded in their actual numbers — no generic advice, no invented figures.

1. Get the data

The Caliper dashboard (this plugin) reconstructs spend from ~/.claude/projects transcripts. Prefer its API:

Find a running server: try curl -s http://localhost:8787/v1/health; if that fails, check other ports the launcher may have picked, or start one with this plugin's scripts/launch-control-tower.mjs (it prints the URL).
Pull, in order of value:
- GET /v1/aggregate — machine-wide totals, by-day / by-repo / by-tier (poll until done: true; the first scan is incremental).
- GET /v1/sessions/all — every session with per-session cost, model tier, wf/sub counts.
- GET /v1/observed + GET /v1/observed/:id — workflow runs with per-call tier/cost/cache telemetry (for the currently selected session).
If no server can run, say so and analyze what the user pastes instead.

2. Analyze — find the levers

Ground each finding in a number you actually fetched. Look for:

Tier-task mismatch: expensive tiers (fable/opus) doing mechanical work (repo exploration, file reads, formatting). Compare per-repo and per-session tier splits.
Delegation gaps: sessions where the main conversation dominates cost — work that could have gone to cheaper subagents/workflow agents with isolated context.
Workflow model mix: runs where every agent() call used the default tier; planning/review can justify opus, implementation usually shouldn't.
Cache economics: cache reads are ~10× cheaper than fresh input, and 1-hour cache writes cost 2× the 5-minute rate. Reward stable system prompts/context; flag patterns that churn the cache.
Repeat spend: the same repo/task shape recurring daily (visible in by-day/by-repo) — candidates for a skill, a workflow, or cached artifacts instead of re-derivation.

3. Recommend, then offer the skill

Present the 3–5 biggest levers with the evidence and an estimated impact each.
Ask whether to write a personalized skill at ~/.claude/skills/cost-discipline/SKILL.md capturing the durable rules (e.g. "delegate repo exploration to haiku subagents", "use opus for planning, sonnet for implementation in workflow specs", "keep CLAUDE.md stable to protect the cache ratio"). Only write it after a yes.
The generated skill must cite the data snapshot (date + headline numbers) it was derived from, so it can be re-audited later.

Honesty requirements (non-negotiable)

Every dollar figure is a cache-aware estimate reconstructed from transcripts, not a billed amount — label them as such.
State sample sizes; never turn one expensive session into a universal rule.
Cheaper models can need more attempts: present substitutions as token-economics ceilings, not promised savings.