cost-optimizer
Trigger when the user asks to audit Claude Code costs, reduce token spend, says "my Claude bill is too high", "optimize my CLAUDE.md", "why is this project burning tokens", or "/cost-optimizer". Scans a project for the common Claude Code cost leaks and returns a prioritized fix list.
Оригинален текст · няма потвърден превод на българскиПърво прочетете инструкцията.
Файлът се зарежда от посочения източник само по ваше действие. Можете да го прочетете тук, да го копирате или да го запазите като .md.
GitHub хранилище- Звезди на репото
- 3977
- Лиценз
- MIT
- Последна проверка на записа
- 21.09.2026 г.
Добавете в своя инструмент.
Командата отваря избора на skills от репозитория. Прочетете файла и условията на автора преди изпълнение.
npx skills add mergisi/awesome-openclaw-agentsСтраницата копира текст. Не инсталира и не изпълнява нищо на устройството ви.
Откъс от SKILL.md
Началото на файла на автора, без превод.
Cost Optimizer
Audit a project's Claude Code setup for the handful of well-known cost leaks and output a prioritized remediation list. This is a static inspection — no pricing API calls, no guesswork about past spend.
When to use
- "My Claude Code bill doubled this month — what's going on?"
- "Audit this project for token waste"
- "Why is my session context always huge?"
- Post-mortem after a runaway agent run
What to check
Run each check and record PASS / WARN / FAIL with a short note.
1. CLAUDE.md size
wc -l CLAUDE.md .claude/CLAUDE.md 2>/dev/null
- PASS: under 100 lines
- WARN: 100-300 lines
- FAIL: >300 lines — this is prepended to every message, so every line costs tokens forever.
Fix: move reference material (API keys, runbooks, verbose examples) into skill files or docs/, and leave CLAUDE.md as a thin pointer.
2. Memory file bloat
find ~/.claude/projects -name 'MEMORY.md' -exec wc -l {} \;
find .claude/memory -type f -exec wc -l {} \; 2>/dev/null
- WARN if any memory file is >500 lines.
- FAIL if >1500 lines — memory should be a curated index with links to detail files, not a dump.
Fix: extract old sections into memory/<topic>.md detail files and link from the index.
3. Unused or shadowed skills
ls ~/.claude/skills/ .claude/skills/ 2>/dev/null
- Cross-reference with the skill invocations in the last 30 days of shell history or session logs if available.
- Skills that have never been invoked but have huge
SKILL.mdbodies (>300 lines) are loaded into the skill index every turn — prune them.
4. Subagent fan-out patterns
Grep the project for Task/agent launch patterns:
grep -rn 'TaskCreate\|subagent\|parallel.*agent' --include='*.md' --include='*.js' --include='*.ts' .
- WARN if the repo spawns >5 parallel subagents per run without aggregation — each spins up its own context window.
- FAIL if subagents re-read the whole repo instead of receiving a scoped prompt.
Fix: pass a short brief + a handful of file paths to each subagent, not the entire CLAUDE.md context.
5. Prompt cache hygiene
Look for .claude/settings.json and check if the project sets an unusually short cache TTL, or if hooks invalidate the cache every message (e.g. a hook that echoes a timestamp into the system prompt).
- FAIL any hook that injects
date/uuidgen/ random content into every turn — it kills the cache, 10x cost.