A meeting notetaker that shows its receipts

pebbl. lives in the menu bar. It records your microphone and, on newer Macs, the call audio as a separate track. It turns speech into text on your Mac, then writes notes where every decision and action links to the exact moment it was said.

Notetakers fail quietly, in three ways. They look like they're recording when they aren't, and you find out afterwards. They write confident summaries you can't check, so one wrong action poisons the rest. And they overwrite last week's decision with this week's, as if it never happened. Many also join your call as a bot, which changes the room before anyone has spoken.

I mapped one meeting and the week after it against the five jobs in the spec: capture, understand, remember, verify and act. The meeting itself isn't where it hurts. The pain sits afterwards, when you need to trust, find or act on what was said. So that's where pebbl. spends its effort.

• Small. No clutter, no performative complexity.
• Dependable. Accurate state, resilient capture, recoverable data.
• Quiet. Useful when needed, invisible when not.
• Grounded. Show the evidence, never invent certainty.
• Actionable. Turn remembered conversations into work.

Each one settled real arguments later. Grounded is why a paraphrase can't become a quote. Dependable is why the word Recording waits for real audio. Quiet is why nothing joins your call.

What: every decision, action and chapter carries a citation: a transcript line and the words quoted from it. Click it and you hear that moment. Why: a summary you can't check is a rumour. The fastest way to lose someone is one confident, wrong action item.

The important call was where the check lives. Asking a model to be honest is a hope; checking its answer is a guarantee. So the model returns citations, and native code decides. A quote found word for word is verified. A real line with different words is shown as AI interpretation. A citation to a line that doesn't exist is dropped, and the count is shown rather than hidden.

What: a list of the things you said you'd do, across every meeting, with due dates. Why: that's the list people actually need on a Friday afternoon, and models are generous with ownership. So the rule is strict and written in code: the model says you own it, the quote is verified, and the line came from your own microphone track.

What: the recording lifecycle lives in the native layer, and the interface only shows what it reports. Why: the worst failure a notetaker has is silent. It only says Recording once the first real audio arrives. Pure silence becomes “No sound coming in” after 1.5 seconds. Audio is written in 30-second chunks with the header rewritten every second, so a crash costs about a second.

What: speech becomes text on the Mac with Whisper. Notes are optional, use your own Anthropic key from the Keychain, and send transcript text only. Why: people are recording their colleagues, so the default has to be that nothing goes anywhere. Separate tracks also mean “you” and “others” come from the track, not from guessing voices.

What: a trail of decisions across meetings, with links like Replaces it, Builds on it and Conflicts with it. Why: plans change, and a tool that quietly overwrites the old answer makes every argument about who said what. pebbl. never rewrites history. It suggests a connection from shared wording, changes nothing until you choose, and keeps an append-only record.

The Trail shipped first and the Map second, on purpose. A graph drawn before the data model works is decoration. Once links and history held up in tests, the Map became a second way into the same facts, with a screen-reader list stating every connection in words.

I wireframed nine screens on the same grid as the final build, so each numbered note sits on the same element in both. Switch between Wireframe and Built to see what changed, and what didn't.

Wireframes · nine screens, one grid

Drawn low-fi first, then built on the same grid. Every number sits on the same element in both views.

WireframeBuilt

Menu barTodayRecordingMeetingWalkthroughSearchDecisions: TrailDecisions: MapSettings

123456

Menu barWhy this screenThe menu bar is the product. Everything you need mid-meeting fits in one popover, and it works with the main window closed.

• Mineral warmth. Chalk and paper surfaces, ink text. Moss means live and healthy, clay means recording, slate-blue means evidence.
• Never colour alone. Every status pairs a colour with a word and a shape. Problems use a square marker.
• Contrast is tested. The muted grey was darkened until it passed WCAG AA on every surface, and a test checks the pairs in both themes.
• Motion has a job. Still when idle, breathing while listening, a slow drift while processing, and static for errors. Reduce Motion is respected.
• Licensed type. Geist, bundled locally under the OFL. Timestamps in the system mono.

I wrote the spec first, and that's where most of the design happened: positioning, jobs to be done, what V1 must do, what it must not pretend to do, and phases with exit gates. Then I built it with Claude in one sitting, logging each decision as it was made.

1. Phase 0Foundation

   Tauri shell, menu bar, popover, design tokens with contrast tests.

2. Phase 1Capture

   Microphone and call audio to crash-safe files, with recovery.

3. Phase 2Evidence

   On-device transcription, checked notes, Evidence Mode, My promises.

4. Phase 3Library

   Search, editing, export, retention, shortcuts.

5. ThenScreen and brand

   Screen recording, the brand system, chapters, the decisions trail.

6. Phase 5Decision graph

   Links, append-only history, suggestions and the Map.

The non-goals mattered as much as the features. V1 refuses to pretend it can capture every meeting automatically, tell every voice apart, be perfectly accurate, or get consent on your behalf. Saying no is what kept a one-day build honest.

An app about not overclaiming shouldn't overclaim about itself. The status file uses one rule: “tested” means verified by a named command in a named environment. Here's where each piece stands.

• The spec was the design work. Writing down the non-goals and exit gates did more for the product than any screen.
• Trust is mostly what you refuse to show. A dropped citation, a hidden echo line and a missing promises list are all features.
• Building in the real material surfaces real decisions. “Recording only after the first buffer” and echo removal never come up in a mockup.