tl;dv

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me the call-clipping tool I actually need instead of tl;dv.

The idea worth copying is narrow and good: **nobody watches a fifty-minute call, and everybody watches a forty-second clip of the customer saying the thing.** Build for the clip.

STACK
- Node 20+ with Fastify, server-rendered HTML
- SQLite through better-sqlite3, WAL mode, FTS5 over transcripts
- ffmpeg for clipping, whisper.cpp for transcription
- Object storage with a delivery network for the media
- Caddy in front

THE DATA MODEL
- calls: id, title, started_at, duration_ms, media_path, participants_json, deal_ref
- segments: id, call_id, speaker, start_ms, end_ms, text
- clips: id, call_id, start_ms, end_ms, title, note, created_by, token, visibility, created_at
- clip_views: id, clip_id, viewer_hash, watched_ms, at — **who watched it and how far is the point of sharing it**
- reels: id, name, clip_ids_json — several clips from different calls in one sequence

CLIPPING, WHICH MUST TAKE FIVE SECONDS
- Select the transcript lines, not a time range. Selecting text and pressing a key is fast; dragging on a timeline is not
- Pad a couple of seconds each side so the clip does not start mid-word
- Render with ffmpeg on a queue, stream copy where the cut lands on a keyframe and re-encode only when it does not
- Burn in captions as an option — **a clip pasted into a chat is watched with the sound off**, and without captions it says nothing

SHARING
- A token URL with an open-graph preview showing the poster frame and the quoted line, because that is what makes somebody click
- Visibility per clip: anybody with the link, or signed in. Default to the narrower one — these are recordings of customers talking
- An expiry option, and a revoke that actually works

THE REEL, WHICH IS THE FEATURE PEOPLE UNDERESTIMATE
- Ten customers saying the same thing, concatenated into two minutes. That is the most persuasive artefact a product team can put in front of anybody
- Assemble by dragging clips; render once and cache

SEARCH ACROSS CALLS
- Full-text over segments, results as moments with a play button and a one-key clip
- Filter by participant, date and deal

CONSENT AND RETENTION
- Announce the recording, record that you did, and keep the archive visible to participants
- A retention sweeper that deletes media and rows, and deletion on request. **A clip is a person's face and voice on a link — treat the revoke as a real requirement, not a checkbox**

WHAT MATTERS MOST
Clipping from the transcript and burnt-in captions. The first makes it fast enough to be used; the second makes the clip work where it is actually pasted.