Can a prompt replace Tactiq?

Build a browser extension that captures live meeting captions and turns them into notes. It must never record audio.

Capture: a content script that activates on the meeting platform's own domain and finds the caption container. Attach a MutationObserver to it. Caption panels do not append clean lines — they re-render the current utterance repeatedly as the recogniser corrects itself, then finalise it. Handle this explicitly: keep the current in-flight utterance per speaker keyed by the caption element's identity, replace it on each mutation, and only commit it to the transcript when it stops changing for a short debounce window or the element is removed. Deduplicating naively by string comparison produces a transcript full of half-sentences, and this is the single hardest part of the build.

Speaker attribution: read the speaker label the caption panel provides. Where it provides none, mark the segment as unattributed rather than guessing.

Selector resilience: keep every DOM selector in one configuration object with a version and a last-verified date, not scattered through the code. Detect a capture failure — observer attached, no captions for a long stretch of an active meeting — and surface a visible warning in the extension rather than producing a silently empty transcript. Say in the README that these selectors will break and that this file is where to fix them.

Session handling: start capture when the user clicks record in the extension, show a persistent indicator while it is running, and stop on click or when the meeting page closes. Never start automatically.

Storage: transcripts in the extension's local storage, exported to a local SQLite-backed companion or downloaded as files. Each transcript keeps the meeting title, date, duration and full timestamped segments.

Notes: on demand, send the transcript to Anthropic or OpenAI, whichever key the user configured, and produce a summary, decisions, open questions and action items with owners where named. Never send automatically; the user presses a button after reading the transcript.

Out of scope: recording audio or video, joining meetings as a bot, real-time note generation during the call, and any hosted backend or account. Note in the README that participants should be told captions are being captured, and that the tool captures only what the platform is already displaying to everyone in the room.