Edit Your Shorts With Claude

---
name: short-form-cut
description: Turns raw talking-head footage or a long landscape recording into finished vertical shorts — moment selection, speaker-tracked reframe, audio mastering, captions, and on-screen hook titles — with the three human gates that keep it from shipping silent failures. Trigger on "cut this into shorts", "make reels from this recording", "edit my video with Claude", "turn this podcast into clips".
---

# Short-Form Cut

Raw footage in, finished captioned vertical shorts out, no editor, no timeline. The
complete pipeline is open-source at github.com/drleadflow/short-form-content-factory —
this skill is the operating procedure that runs it (or any equivalent tool chain), and
more importantly the three gates where a human has to look.

## Before you start (connector-first)

You need the footage on disk or in connected storage (Drive/Dropbox via a connector), a
transcription route (Whisper locally, Deepgram, or a connected transcription service),
and a caption/render tool. If the repo is installed, everything below maps to its
stages. Manual fallback: any editor — the gates and the rules still apply.

Ask up front: platform(s), target length (45–75s is the sweet spot from our data), and
whether the footage is already captioned (double-captioning ruins a whole batch).

## The pipeline

**1. Dedupe.** Byte-identical files, filename variants, AND alternate takes (filenames
never reveal these — compare transcripts). A 27-file batch is usually 16 real clips.

**2. Prep.** Extract audio twice: high quality for mastering, 16kHz mono for transcription.

**3. Transcribe.** A base model is enough to CHOOSE moments; use a bigger one only when
the words become captions.

**GATE 1 — a human reads the transcripts and picks the moments.** "Which 40 seconds is
worth posting" is a judgment call. This is also where transcription errors get caught
before they're burned in.

**4. Cut + reframe.** For a landscape multicam master, the reframe must be
speaker-tracked, not a center crop: detect shot boundaries, then center the 9:16 window
on the moving subject (motion = whoever is talking). Low-confidence shots get flagged
and the render refuses until they're resolved — a flag beats a silently mis-framed clip.

**5. Enhance + blend.** Never ship AI noise-removal at full strength — it scrubs the room
out and speech reads synthetic. Blend enhanced over original: ~80% clean indoor, ~65%
moderate room noise, ~50% wind or overlapping speakers. Target −14 LUFS, peaks under
−1 dBTP.

**GATE 2 — listen.** Render two blend ratios at once and pick by ear.

**6. Captions.** Check a mid-clip frame FIRST for existing captions. Word-level,
readable, safe-zone aware for the platform's UI overlays.

**7. Title hooks.** Two lines, 2–4 words each, all caps, first three words carry the
tension. **A title that labels the topic is dead** — if a viewer can read it and nod
"makes sense," rewrite it to withhold the payoff. Color-code the second line by hook
type (warning / how-to / reframe / opportunity). Hold ~3.5s, fade out.

**8. Composite + QA.** Pair titles to clips by filename stem. Then build a contact sheet.

**GATE 3 — look at the contact sheet.** Every failure this kind of pipeline produces is
silent: a wrapped title, a transparent title, a title colliding with captions, audio
normalized into clipping, a batch that "shipped" at −40 LUFS and was inaudible on a
phone. None return an error code. "It ran" is not "it's right."

## Deliverable

For each short: the file, its hook title, a caption with a comment keyword, and the
loudness reading. Plus the list of moments that were great ideas but weak deliveries —
re-film prompts for next time.

## Publishing

Load into the scheduler (Blotato or similar via connector) as drafts. Confirm with the
human before anything publishes. Always.