claude-roast

The opposite of Claude Code Insights. Reads your ~/.claude history and tells you, in detail, how badly you use Claude Code.

npx claude-roast

No install. No API key. Reads only your local history, then shells out to your existing claude CLI (claude -p) to write the actual roast — using whatever auth you've already set up for Claude Code. Only a stats summary is sent (no prompt content).

What it measures

Mined from ~/.claude/history.jsonl and session transcripts:

| Category | What it catches | | --- | --- | | Prompt habits | average length, one-word prompts, pasted blob count | | Attitude | corrections (no, wrong, actually), lazy fix it prompts, profanity, all-caps rants, apologies (yes, to the AI) | | Overprompting | ultrathink / think harder instances, IMPORTANT shouts, very very pile-ons | | Danger zone | --no-verify, git reset --hard, git push --force, rm -rf | | Time patterns | late-night sessions, weekend prompts, hourly histogram, longest streak | | Slash commands | what you actually reach for | | Top projects | where the prompts go | | Tool pain | tool failures, permission denials |

Then it ships the numbers (not the prompts) to an LLM and asks for a roast.

Usage

# Default: mean severity, all history
npx claude-roast

# Severity dial
npx claude-roast --severity gentle
npx claude-roast --severity mean      # default
npx claude-roast --severity nuclear

# Scope
npx claude-roast --days 30
npx claude-roast --project iqai-prediction

# Just stats, no model call
npx claude-roast --dry-run

# Generate anti-CLAUDE.md to ./CLAUDE.roast.md
npx claude-roast --rules

# Generate a shareable HTML report (default ./roast.html)
npx claude-roast --html
npx claude-roast --html ./out/roast.html

# Machine-readable
npx claude-roast --json

# Override the model passed to `claude -p`
npx claude-roast --model claude-sonnet-4-6

Requirements

You need the claude CLI on your PATH. That's it — no API keys, no config file. claude-roast shells out to claude -p and uses whatever auth you've already set up for Claude Code.

--dry-run, --json, --rules, and --html (without a roast) never spawn claude and work without it installed.

HTML report

--html writes a self-contained, dark-themed page with the roast, a stats grid, an SVG hour histogram, top projects, and your favorite slash command. Single file, no external CSS or JS — open it, share it, screenshot it.

Privacy

• Everything runs locally.
• The roast call sends aggregated counters and short samples (5 one-word prompts, 5 corrections, etc.) — never your full prompt content or code.
• --dry-run, --json, --rules, and --html (without a roast) never hit the network at all.

How it's built

• TypeScript, ESM, Node ≥ 20
• claude -p — the roast call shells out to your local Claude Code CLI; no API keys, no SDK
• Zod — single source of truth for schemas (history entries, session messages, CLI options, metrics)
• picocolors + yocto-spinner — modern, tiny terminal output (e18e)
• commander — CLI parsing
• tsup — single-file ESM bundle

The metrics layer is split into per-category extractors (collectPromptHabits, collectAttitude, collectDanger, …) so it's easy to add new sins. PRs welcome.

Inspired by

@aidenybai's Claude Doctor — same shape (read ~/.claude, point out the patterns), opposite tone (roast, not therapy).

License

MIT