reel-editor-critique

SKILL.md

---
name: reel-editor-critique
description: >
  Critique a finished @ImplexaAI reel (and its script) the way a top-1% short-form
  video editor would — brutally and specifically. Catches AI-sounding delivery
  (letter-by-letter reads like "D.M."/"U.R.L.", robotic abbreviations, awkward
  spoken seams), clumsy phrasing, weak hooks, pacing/dead-air, caption-sync and
  overlay problems, AND content gaps (a key message the reel should make but
  doesn't — e.g. "you can build/deploy this with a prompt in Claude Code or
  Codex"). Use as the final QC gate before posting any reel, when someone says
  "critique this reel / video like a real editor", "what's weak about this",
  "does the script land", "why does this sound AI", or to review a reel script
  before the avatar is rendered.
---

# Reel editor critique — the top-1% editor's eye

Most reel critiques rubber-stamp ("looks good, ship it"). This one does the opposite:
assume the reel is NOT good enough yet and find every reason a sharp editor would
send it back. Be specific, cite timestamps and exact lines, and give the fix — not
"tighten the hook" but the rewritten hook. Two passes: the **script** (can run before
the avatar exists) and the **finished video** (after render).

## Cost principle — regeneration is expensive, so the script pass is where this skill earns its keep
A HeyGen avatar re-render costs real credits and minutes, and a re-render forces a new
Whisper pass + re-pinning every overlay. So the goal is **zero re-renders: get the
script and the visual plan perfect BEFORE the avatar is generated.** Run the full
critique on the SCRIPT + plan at the approval gate and hold nothing back there — every
AI-tell, content gap, awkward seam, and pronunciation risk you catch pre-render is one
you don't pay to fix later. Finding problems only in the finished video is a failure
mode of the process, not a success of the critique. Push this discipline upstream too:
the debate/brief steps should pressure-test the script for the brand must-says (e.g.
"build/deploy by prompting Claude Code or Codex"), letter-by-letter abbreviations, and
clumsy spoken seams, and lock a clean script before anything renders. Treat a HOLD
discovered post-render as a lesson to tighten the pre-render pass, not as normal.

## Inputs to load
- **Script:** `REEL_BRIEF.md` / `HEYGEN_BRIEF.md` (the intended words).
- **Actual transcript + timings:** `src/subtitles-words.json` (whisper word-level). This is what was ACTUALLY said — diff it against the script and read it aloud in your head to catch AI-tells and awkward seams at real timestamps.
- **The video:** extract frames from `out/day-N.mp4` (hook, each overlay/card beat, the close, the cover) for the visual pass. Check `volumedetect` for audio level/dead air.
- **Brand key-messages checklist:** `references/brand-key-messages.md` — what an Implexa dev-series reel must convey. Gaps against this list are findings.

## The six dimensions (score each, then list findings)

1. **Hook (first ~3s).** Does it stop the scroll and promise something? Is the first spoken line + first caption doing work, or warming up? A vague or throat-clearing open is a finding.
2. **Script & message.** Is the core point unmistakable? Does every sentence earn its place? Is the logic clean? **Content gaps** — run the brand-key-messages checklist; a missing must-say (e.g. *you can do this with a prompt in Claude Code / Codex*) is a HIGH finding, because that's the differentiator the audience came for.
3. **Voice & delivery (the AI-tell pass).** Read the transcript as speech. Flag, with timestamps:
   - **Letter-by-letter / abbreviation reads** — "D.M.", "U.R.L.", "A.P.I." spoken as letters sounds robotic. Rewrite to natural speech ("send me the link", "your live link").
   - **Awkward seams / list-to-sentence collisions** — e.g. "…git, deploying. today you actually ship" — a list that dead-ends then hard-cuts. Smooth it.
   - **Run-ons, tongue-twisters, words the TTS mangles** (brand names like "Supabase"/"Vercel" — confirm pronunciation; if risky, respell phonetically in the script).
   - **AI cadence** — "not X, it's Y" antithesis, "here's the thing", over-even sentence lengths, em-dash overload. Spoken or written, flag it.
4. **Pacing & retention.** Dead air, a beat that drags, the mid-reel drop-off point, whether the energy builds. Is there a 2-3s stretch where nothing new happens? Mark it.
5. **Visual & editing.** Caption sync to the audio, overlay/card legibility + collisions with the face, whether the rebuilt app surfaces land on their beat, brand consistency, and the **cover frame** (clean expression? readable promise?).
6. **CTA & loop.** Is the ask concrete and singular? Does the close earn the follow / drive the comment? Does it set up the next episode?

## AI-tell quick list (auto-flag any of these)
Letter-spelled abbreviations read aloud (D.M., U.R.L., A.P.I., F.A.Q.) · "delve", "in today's fast-paced world", "it's not just X, it's Y" · signpost phrases ("here's the thing", "let's dive in") · uniform sentence length · em-dash every sentence · a list that ends mid-air then hard-cuts · brand names the TTS is likely to mispronounce.

## Output contract — write to `CRITIQUE.md`
1. **Verdict:** SHIP / HOLD / RESHOOT, one line, with the single biggest reason.
2. **Scorecard:** the six dimensions, each 1–5 + a one-line why.
3. **Findings table:** `# | severity (blocker/high/med/low) | dimension | timestamp/line | the problem | the FIX (exact rewrite or re-time)`. Order by severity. Be specific enough that someone can apply the fix without rewatching.
4. **Rewrite block:** if the script needs changes, supply the full corrected script (spoken-voice), so the next avatar render is one copy-paste.
5. **What's strong (keep):** 2-3 things not to touch.

## Decision
- Any **blocker** or ≥2 **high** findings → HOLD/RESHOOT; the script-level fixes usually mean a re-render of the avatar, so resolve them before HeyGen, not after.
- Be willing to say RESHOOT. A polished reel that misses the differentiator (e.g. never mentions building with a prompt) is worse than a rough one that nails it.

## References
- `references/brand-key-messages.md` — the must-convey messages + AI-tell patterns specific to the @ImplexaAI dev series.
- Pairs with the `red-team` step (which stress-tests the angle/plan); this one judges the finished CRAFT — script, delivery, edit.