tokenwatch
Skill1 starsChanged 38 days agoMITView on GitHub
Report Claude Code token burn and cost, and say what to change next. Use whenever the user asks how we are doing on tokens, cost, spend, or burn - "how are we doing", "how's our token usage", "are we saving anything", "check spend", "is the autocompact working", "did offload fire". Also use before and after changing any cost-related setting (autoCompactWindow, effort, model, CLAUDE_CODE_SUBAGENT_MODEL) so the change is measured rather than assumed.
In this repository5 other files it gives its agents.
- skills/fable-setup/templates/CLAUDE.mdCLAUDE.md
- delegateSkill
- fable-setupSkill
- product-videoSkill
- wrapSkill
The file
Files
Token usage check
Measures what actually happened from Claude Code's own transcripts. Answer first, then recommend. Do not make up figures: every number comes from the script.
Run it
python <path-to>/token-usage-report.py --days 7--days 1 for a same-day check after a settings change. --project <name> to limit it to
one project folder. --no-log to skip appending to token-usage-history.csv. Every run
without --no-log appends rows, and that CSV is the only lasting record: Claude Code deletes
transcripts after cleanupPeriodDays (30 by default).
For trend, read token-usage-history.csv and compare usd_per_turn by run_date.
$/turn is the metric that matters, not total spend. Total moves with how much work got
done that week; $/turn shows whether the configuration is working.
Reading the result
Diagnose from the cost composition, then name one change.
- cache_read is the largest line and session ctx_avg is high - context is growing
unchecked, and every turn re-reads it. Set
autoCompactWindowin~/.claude/settings.json(try 300000, then 200000) and use/clearbetween unrelated tasks./clearcosts nothing;/compactis itself a large request. - cache_write is high - new context is written to cache at a premium (1-hour writes cost
2x input) and then re-read every turn. Split the cause before acting:
- big rewrites with no idle gap are compaction: the window is too tight; raising it trades fewer rewrites for more re-reads, so measure both;
- big rewrites after an hour or more idle are cache expiry:
/clearor finish before walking away from a large session; - steady per-turn growth is tool output: trim with
head,--limit, targeted reads, or hand bulky work to a subagent.
- output above ~25% - thinking and text. Lower effort for routine work (
/effort). - subagent turns near 0 while work had separable bulk - offload is not firing. Check
that
CLAUDE_CODE_SUBAGENT_MODELis set (otherwise subagents inherit the main model and save nothing) and thatCLAUDE.mdtells Claude when to delegate. - expensive-model subagent turns - usually a fork (forks inherit the parent model) or an agent definition that pins a model. Name the sessions and ask whether it was intended.
When a model is missing
The warning names any model absent from the PRICES table in the script; those turns are
left out of totals. Add current rates from the claude-api skill or Anthropic's pricing
page, not from memory.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

