The Lab Notes prompts: a local-first app with Claude Fable 5.2

Lab Notes is a local-first web app for logging AI experiments: what you tried, the exact prompt, the input and output of each run, and what worked and what didn’t. It came from one build prompt plus one short resume prompt, run with Claude Fable 5.2 in Claude Code. Both prompts and the test record are free to download. Get the files

The app’s production build, recorded in Chrome at 1440×900. The experiments are demo data the app generates.

What the prompts do

PROMPT-1.txt asks for a local-first app in Vite and TypeScript with no backend, no accounts and no API keys; all data stays in the browser (IndexedDB). It covers experiments with runs (exact prompt, input, output, images, time, cost and a pass/fail note), a side-by-side prompt diff, a dashboard with search, filters and charts drawn without a chart library, Markdown, ZIP and JSON export, keyboard shortcuts, light and dark mode, a 500-experiment demo button, Vitest tests and an honest README.
PROMPT-2-resume.txt is the short message sent in the same session after the first run hit the 5-hour usage limit: finish the original brief, make the tests pass, write the README, and don’t start over.
README.md and TESTED.md (in the zip) say how to reproduce the run and what was checked in a browser: what passed and what didn’t.

How to use them

Make an empty folder and open Claude Code in it. I used Claude Fable 5.2 (claude-fable-5-2) in Claude Code 2.1.322 on a MacBook Air M1, and you need Node and npm.
Paste PROMPT-1.txt as the first message and let it run.
If the session stops at a usage limit, wait for the reset and paste PROMPT-2-resume.txt into the same session.
When it finishes, run npm install, npm run build, npm test and npm run preview, then open the local address it prints.

Your result will not match mine exactly. Two installed Claude Code skills shaped my run: the bundled dataviz skill picked the chart colours, and the hallmark design skill picked the look (“Console” theme). Without them you will get a different design.

What didn’t work

Verdict: PASS with caveats. The build is clean and all 80 Vitest tests pass. In the browser, experiments and runs, the prompt diff, the ZIP and JSON exports, the shortcuts, both themes and the 500-experiment demo worked.

Major: on a 390 px phone the open run panel is clipped on the right, and “Delete run” can’t be reached.

It took two prompts, because the first session hit the 5-hour limit before the stylesheet was written, and Claude never opened the app in a real browser itself. It hasn’t been tested on a real iPhone, in Safari or in Firefox. The minor issues are listed in TESTED.md.

The files

The zip holds both prompts, README.md and TESTED.md. The app’s code isn’t included. Free, no sign-up.

Built by Claude Fable 5.2 in Claude Code. The look comes from the hallmark skill and the chart colours from the dataviz skill. I run AI experiments and write them down: what worked, and what didn’t.

All free guides on this site