promptstreak

Your Claude Code prompting habits as a GitHub-style contribution graph.

Demo: promptstreak.pages.dev · get yours: npx promptstreak

A commit graph stopped meaning anything the moment agents started committing daily. This one counts the part you actually did: the prompts you wrote. Ten metrics, one year, one screenshot.

npx promptstreak

Reads your local Claude Code history, prints the graph in your terminal, and opens a shareable version at promptstreak.pages.dev. Nothing is uploaded — not by the CLI, and not by the page.

The grid switches between fifteen metrics. Four of them are the interesting ones:

Plus prompt length, spec-shaped prompts, overnight handoffs, prompts at night, politeness, the plain prompt count, and a mood family — swearing, annoyed, caps lock, thanks, emoji — that also works as an overlay on any of the others.

npx promptstreak                          # graph + stats, opens the shareable page
npx promptstreak --metric leash           # pick the metric that fills the grid
npx promptstreak --print-url              # print the share link instead of opening it
npx promptstreak --no-open                # terminal only
npx promptstreak --json                   # raw metrics, for scripting
npx promptstreak --local                  # history.jsonl only, skip transcripts
npx promptstreak --no-quotes              # keep "Things you said" out of the link
npx promptstreak --curate                 # let your local `claude` pick the quotes

Prompts on a server or a second laptop live in that machine's history. Merge them:

ssh box 'cat ~/.claude/history.jsonl' > box.jsonl
npx promptstreak box.jsonl

Any number of files or directories can be passed, and duplicates are dropped. The web page takes dropped files the same way.

The run ends with up to ten of your own lines — the swearing, the caps lock, the apologies — and they travel with the link so the page can show them. They are redacted first (links, paths, emails and anything token-shaped are stripped) and printed in the terminal before the browser opens, so you see exactly what ships. --no-quotes keeps them home. --curate hands the top candidates to your local claude CLI to pick the funniest, and uses the built-in ranking if that fails.

On the page, select any words in a quote to blank them out, or remove a quote you would rather not show; the link in the address bar updates as you go, so what you share is what you see. Only your browser keeps the originals, so only there can they be undone or restored.

Two files, with very different reach:

• ~/.claude/history.jsonl — every prompt you submitted, going back roughly a year. This is what fills the grid.
• ~/.claude/projects/**/*.jsonl — full session transcripts, but Claude Code prunes these after cleanupPeriodDays (about a month by default). This is the only place the agent's side of the conversation lives, so autonomy, the longest unattended run and token usage come from here, and they cover only the days that survive.

If you want a fuller picture next year, raise cleanupPeriodDays in your Claude Code settings now. Nothing can recover transcripts that have already been pruned.

$CLAUDE_CONFIG_DIR and $XDG_CONFIG_HOME are both honoured, and every directory that exists is merged rather than the first one winning.

Nothing you typed leaves your machine unless you let it. The one thing that can is the quotes, and you see them before they go.

• npx promptstreak reads ~/.claude/history.jsonl and the session transcripts on disk, counts them locally, and prints the result. Its only network action is opening a browser tab; it sends no telemetry and calls no API.
• The link it opens carries view data, never the source — per-day counts, headline numbers and, unless you pass --no-quotes, the redacted quotes printed in the terminal — compressed into the URL fragment, which browsers never send to a server. File paths and project names are not in it.
• The page is static: no backend, no upload endpoint. Dropped files are parsed in the browser and stay there. Saved machines live in your browser's local storage. Google Analytics counts visits; it is configured to receive the page address without the fragment, so the share data never reaches it.

Read the source if you want to check: src/parse-core.ts is the whole of what is extracted, and src/mood.ts is everything that touches prompt text.

npm install
npm test          # fixture-based; every metric is asserted against hand-computed values
npm run build     # CLI to lib/, browser bundle to web/app.js
npm run dev       # run the CLI from source
npm run web:dev   # serve the page locally

Tests run under TZ=UTC because every metric buckets by local calendar day — a night owl's day must not split at midnight UTC.

MIT © Nikita Savchenko