clippr — what it is, its parts, where we are

what it does, in one breath

give clippr an episode or a film; it reads the subtitles, finds the moments that work on their own, and cuts each one into a vertical clip under 20 seconds, with or without music, ranked for you to review and download.

• for people scrolling shorts who have never seen the show — every clip must make sense alone
• two shapes of clip: a mini narrative (setup → turn → payoff, joined from two moments if needed) or an engaging situation (a confrontation, a confession, a reveal, a joke that lands)
• nothing leaves the machine: the chosen file is copied in only for the run (removed after), the clips land in a local folder the page points to

the parts (one box each on the board)

flowchart LR
  A[video on this computer
+ .srt or embedded subs] --> B[ingest
clean lines with times]
  B --> C[story finder
4 model passes]
  C --> D[cutter
9:16 · ≤20s · levelled]
  D --> E[music
CC track, ducked]
  D --> F[(output/<episode>/
dry + music + clips.json)]
  E --> F
  F --> G[the page
ranked cards · play · download · open folder]
  A -.-> G
  

how the AI decides (the four passes)

1. read the episode — characters and their relationships, scenes with what is physically happening, and "threads": setups that might pay off later (a lie, a threat, a running joke).
2. propose the moments — every candidate that passes the stranger test ("cover the rest of the episode, read only these lines: is it clear who wants what, and does it end on something?"), with a score 1–10, a hook (the first two seconds), and a line on why it works alone. Rules: 1–6 segments, ≤20s total, cut on subtitle lines only, a segment is a few lines, never a whole scene.
3. tighten — anything over 20s once measured on real timings is handed back with the exact length of every line and a running total; up to two rounds.
4. the judge — a fresh conversation that sees each candidate ONLY as its own lines, as a scrolling stranger would, and scores understand and interest 1–10. Under 6 on either → dropped, with the reason kept. The card's score is this cold-read score; the author's score is kept alongside.

sequenceDiagram
  participant P as clippr
  participant M as claude-haiku-4-5
  P->>M: full transcript → "who is who, scenes, threads"
  M-->>P: characters + scenes
  P->>M: "every clip-worthy moment; stranger test; score, hook, why-alone"
  M-->>P: candidate stories
  P->>P: snap to real lines · merge touching segments · measure
  P->>M: "these are over 20s — here is every line's length; tighten or drop" (×≤2)
  M-->>P: tightened stories
  P->>M: fresh chat: "you never saw this show; here are the words of each clip — understand? interest?"
  M-->>P: verdicts → drop < 6, rank by cold-read score
  

how one clip gets made

1. each segment: start 0.3s early, end 0.3s late (the padding shrinks automatically when the story is already near 20s)
2. crop the middle 9:16 of the frame → 1080×1920 at 30 fps (subtitles are burned in only if you tick the box)
3. segments joined with a 0.2s crossfade (picture and sound)
4. dialogue levelled to −16 LUFS — streaming masters downmixed from 5.1 arrive near-silent otherwise
5. a thumbnail at 0.5s, then the music version: track for the mood, looped/trimmed, faded, ducked under the speech, mixed

where the files live

output/<episode>/
  01_dry.mp4  01_music.mp4  01_thumb.jpg   ← ranked, 01 = best
  02_…
  clips.json                                ← everything the page shows (title, scores, judge, segments, files, music credit)
  <episode>.eng.srt                         ← the subtitles that were used (extracted if they were embedded)
  story/  pass1.json pass2.json pass3*.json pass4.json stories.json provider.json prompt_*.md
          ↑ every model answer, for prompt tuning: what was proposed, what was dropped and why
uploads/<episode>/   ← only for uploaded files; deleted after the run unless "keep the uploaded copy"
music/<mood>/        ← the free tracks (tense funny sad heartwarming epic chill neutral) + tracks.json credits

how to run

pip install -r requirements.txt
cp .env.example .env                 # ANTHROPIC_API_KEY
python tools/fetch_music.py --per-mood 3
start_clippr.cmd                     # Windows: double-click → starts the page + opens the browser
uvicorn web.app:app                  # or by hand → http://127.0.0.1:8000
python run.py --video "C:\path\episode.mkv"          # same thing from a terminal; add --srt if there is no embedded track
python tools/make_sample.py          # a synthetic 3-min demo episode if you have no real one at hand

where we are (2026-09-03, end of day 1)

verified (we watched it work)

measured

known limits / not built

decisions to talk about

• quantity vs. quality: the judge threshold is 6/10 on both understand and interest; lower it (CLIPPR_JUDGE_MIN) to ship more, raise it to ship only the best
• padding 0.3s each side vs. total; music bed level (0.35 before ducking reads as "present but subtle")
• music licences: openverse pulls cc0 / by / by-sa / by-nd; by-nd under a video is a derivative → cc0 + by only, or a paid catalogue, before selling
• uploaded copies are deleted after the run by default (files are large and already on disk); the checkbox keeps them
• how do people run this (2026-09-04): browser-only users need a server; the free Render tier only demos it. Next step proposed: E2B sandboxes do the processing per run (credits available), the browser uploads straight to the sandbox, Render only serves the page. Alternative for power users: a packaged desktop app (python + ffmpeg in one zip). Browser-side ffmpeg.wasm rejected (single-threaded, 2 GB ceiling, weak HEVC).
• git: jacesabr/clippr on GitHub (private), Render auto-deploys main; every push is a deploy, so pushes stay explicit