lazysketch

Slides sketched live off your voice. You talk; a small, fast LLM listens to your live transcript and draws the deck while you speak — hand-sketched diagrams, bullets, code blocks, the occasional rubber-duck stamp — on lamp-lit paper, like an overhead projector that grew an opinion.

Try it: https://lazysketch.mortenolsen.pro

The agent drew this one itself, live, off a scripted run of exactly the talk you'd expect — duck requested by voice.

Fair warning: this is a quirky side project built for the wow effect of a deck drawing itself mid-talk. If you want AI to help you make actual presentations, there is real software for that. If you want to say "watch this" and then just… talk — welcome.

What you need

• A Chromium-ish browser (Chrome/Edge/Firefox). Speech models (Whisper for ears, Kokoro for a voice) run on-device in workers — a laptop with WebGPU makes them happy.
• An OpenAI-compatible endpoint of your own (OpenAI, Groq, Ollama, a LiteLLM proxy, …). Nothing is baked in; your key never leaves the browser except to the endpoint you name. A Haiku-class model is the sweet spot — small and fast beats big and slow here.

Everything is client-side: no backend, no account, your presentations live in your browser (IndexedDB), with export/import if you want to move them.

Quick start

1. Open the app → Settings → set base URL + model (+ key), hit Save & preload models.
2. Start a fresh sheet, paste in a talk outline as a cue file — it helps the model far more than you'd expect.
3. Go live, open the mic, and start talking. The deck follows. When you stop, the session is recorded — replay it, scrub it, or resume for Q&A.

No mic handy? Rehearse from a script feeds prepared text through the exact same pipeline, at reading pace or as fast as the agent can draw.

On stage for real

Everyone who presents knows the moment: you're tired, under-rehearsed, you've delivered your points out of order, and the script in front of you has stopped being true. Reading it back doesn't help — what you need is someone who heard what you actually said and can tell you where to go from here. That's the prompter.

• Present opens the slides alone in a clean tab — move it to the projector, double-click for fullscreen. Your original tab keeps the tray, transcript, and gauges: that's your console now.
• Write the talk in the Script section on the desk — full prose or just your talking points — then flip on the Prompter next to Go live. A second loop listens along and keeps the next thing to say written large on your console, streaming in as you speak. On script, it's your own words. Off script, it's a live rewrite from where you actually are: points you've already made are dropped, the ones you haven't are lined up in an order that repairs the talk.
• So when you're lost: take a breath. The line is already there. Say it, and you're back.
• Audience question? It proposes a short answer. Wandered into a story? It holds the bridge home. Timing notes in the script are read, not parsed — it reasons about your pace against the clock.
• The prompter can run on a cheaper, faster model than the artist (Settings), Rehearse runs with it on so you can audition it before trusting it with a real room — and Replay shows its console alongside the talk afterwards, so you can judge every line it offered.

The CLI (optional, unpublished)

There's a lazysketch CLI so a coding agent (or you) can prime presentations from the terminal: author the sheet and cue files as plain files, sync them into the browser, run a scripted test session there — using the browser's LLM settings, credentials never touch the CLI — and get an eval report of what the agent drew.

It isn't on a registry yet, so it's clone-and-link:

git clone https://code.olsen.cloud/incubator/lazysketch.git && cd lazysketch
mise install     # Node 24 + task (or bring your own)
task install && task build
cd apps/cli && pnpm link --global
lazysketch --help

Then in the app: Settings → CLI access → Enable, run the pairing command it shows you once, and you're connected (lazysketch status). The loop is init/pull → edit files → push → run --script talk.txt → report → open --replay. Agents can learn it from the skill in skills/lazysketch. (Safari blocks the localhost bridge; use Chrome or Firefox for this part.)

Honest small print

• Latency is the whole game: with a good endpoint the deck trails you by a few seconds (there's a lag gauge — grease pencil, of course). With a slow model it trails you by an awkward silence.
• The agent sometimes draws the wrong thing, rewrites a block five times, or celebrates prematurely with confetti. This is considered part of the charm.
• The prompter follows meaning, not words — in testing it rode out a cat-on-the-keyboard digression without flinching — but it proposes, you decide. It's a prompter, not a co-author. Occasionally it re-offers the line you're mid-way through saying; it catches up on your next breath.
• Curious how it works? docs/design.md covers the transcript pipeline, the event-sourced deck, and the CLI bridge; DESIGN.md is the paper-world design system.