claw-curses
v1.0.3
Published
Track your AI session frustration over time — curse-word rate and interruption count as proxies for AI performance degradation
Downloads
376
Maintainers
Readme
claw-curses
Track your AI session frustration over time. Reads your local ~/.claude/ session data and plots two proxies for "the AI is annoying me":
| Metric | Source | What it means |
|--------|--------|---------------|
| Curse-word rate | ~/.claude/history.jsonl | How often you swear per message |
| Interruptions | ~/.claude/projects/**/*.jsonl | How often you cancelled the AI mid-task |
Outputs a self-contained report.html in your current directory with daily/weekly charts, a smoothing slider, and a volume-weighted trend line.
Usage
bunx claw-curses # preferred
npx claw-curses # also works (requires Node ≥ 18)Or run from source:
bun src/index.tsThe report is written to ./report.html. Open it in any browser.
How it works
Curse words
Every entry in ~/.claude/history.jsonl is scanned with word-boundary regexes. Repeated characters are handled (f+u+c+k+ catches fuuuck). Matching is case-insensitive. The full pattern list is visible in the Detection methodology section at the bottom of the report.
Days with fewer than 50 messages are shown as gaps in the daily chart — not enough volume to produce a meaningful rate.
Interruptions
Conversation files in ~/.claude/projects/ are parsed as message trees using the parentUuid field. A real human message whose parent is an assistant message with stop_reason !== "end_turn" is an interruption — meaning you sent a new prompt while the AI was still mid-tool-use.
Tool-result messages and system-injected context are excluded.
Smoothing
The daily chart has a slider (default 7 days) that applies a trailing moving average over the last N non-null days. Dragging to 1 shows raw data with gaps.
Trend
The header trend compares the first half vs second half of the window using a volume-weighted rate (Σcurses / Σmessages per half). This prevents low-volume weeks from distorting the signal.
Timeline trim
The timeline starts from the first week that hit 250 messages (5 × the 50-message daily threshold), cutting off the sparse early history.
Data privacy
Everything runs locally. No data leaves your machine. ~/.claude/ is read-only.
Publishing
bun run build # compiles src/index.ts → dist/cli.js
npm publish # runs build automatically via prepublishOnly