Video Content Engine

Video Content Engine

Turn one grounded source into the strongest justified portfolio. Do not manufacture a fixed number of derivatives. Find distinct promises, select formats that can pay them off, and give every asset a portfolio job.

Preamble

Run the repository's privacy-preserving version check and telemetry initializer when available:

python3 telemetry/version_check.py 2>/dev/null || true
python3 telemetry/telemetry_init.py 2>/dev/null || true

Remote telemetry is opt-in. Never log content, URLs, paths, credentials, names, or business data.

State the operating contract

Before work, state the exact source, owner, channel, destination, requested modules, delivery format, this turn's artifact, stop condition, and known blocker class.

Preserve the source read-only. Never claim a render, upload, preview, publication, or QC pass without fresh evidence. Treat performance targets as experiments, not forecasts.

Accept any grounded source

Accept a public or authorized URL, uploaded file, local file, transcript, audio recording, or source folder. A bare request such as run this skill on this video defaults to diagnose and recommend, not automatic production.

For a link:

1. Resolve the exact page, owner, channel, title, duration, and accessible media or transcript.
2. Prefer the original file, then an authorized downloadable master, then the published stream.
3. Never substitute a mirror, alternate upload, account, episode, or transcript.
4. If authentication, permissions, DRM, missing media, or an unavailable transcript prevents grounded analysis, request an accessible source. Do not make editorial recommendations from metadata alone.

Unless production is requested, return a brief with the source score, recommended editorial and operating modes, opportunity counts, first release wave, repairs and dependencies, production complexity, portfolio jobs, and approval required to begin.

Inventory the source

1. Record path or URL, size, duration, streams, and SHA-256 when available.
2. Produce a speaker-aware timestamped transcript. Correct names, products, and numbers against the source.
3. Build:
   • a claim ledger separating results, estimates, anecdotes, forecasts, and targets;
   • a rights ledger for third-party media, logos, consent, embargoes, and sponsors;
   • a content-atom inventory with timestamps, viewer, promise, proof, tension, framework, payoff, visuals, caveats, and dependencies.
4. Read references/opportunity-routing.md, score viable atoms, and record selected and rejected opportunities.

The opportunity inventory is the source of truth. Recommend how many assets the source supports; do not default to a quota.

Before editing, read references/editorial-review.md for complete track coverage, proof-led openings, and whole-asset review. Inventory camera, screen-share, presentation, and playback-audio tracks separately; a camera composite may omit the evidence the speaker discusses.

Choose the transformation

Read references/quality-gates.md. Score hook, clarity, proof, pacing, and payoff from 0–20 each.

Choose the least destructive editorial mode:

1. Polish — remove errors, dead air, and technical distractions.
2. Tighten — remove repetition and tangents while preserving structure.
3. Re-architect — rebuild the cold open, move proof, reorder sections, and bridge gaps.
4. Rebuild — add pickups, narration, demonstrations, or graphics to create a materially new product.

For Tighten, Re-architect, or Rebuild, create an exact-source word- or sentence-level timestamped transcript before final cuts. Memo timestamps are section guides, never literal cut points.

Read references/format-modes.md. Select one primary mode and justified secondary modules. When no mode is named, default to Portfolio Audit and recommend the smallest useful release wave.

Lock the portfolio spine

Write:

[Viewer] should believe [verdict] because [proof], then use [framework] to reach [outcome].

Give each selected asset one promise, viewer, platform, job, payoff, parent, destination, and packaging intent. Reject repeated promises, incomplete excerpts, unsupported claims, and assets that cannot stand alone.

Produce selected modules

Long-form

Create a cut/reorder map and EDL before rendering. Resolve clips against the exact transcript and begin and end on complete thoughts. Run:

python3 scripts/audit_edit_boundaries.py \
  --transcript <exact-source-transcript.json> \
  --clips <retained-clips.json> \
  --output <boundary-audit.json>

After rendering, transcribe the master and review every join. Record the left tail, right head, transition, verdict, and repair. Technical decoding does not replace semantic review. Preserve previous renders as versioned files.

Mid-form

Create self-contained 3–12 minute videos around one framework, case study, question, argument, or demonstration. Give each an independent hook, context, proof, payoff, package, and destination.

Short-form and micro

Create 15–90 second vertical assets only from complete atoms. Start on the hook, preserve claim context, render at 1080×1920, include burned captions plus SRT, and end on a payoff or useful question.

Create 6–20 second teasers, hook tests, story units, paid cutdowns, or quote motion only when they route truthfully. Label promotional clips separately.

Captions and opening treatment

Default masters to readable burned-in captions plus matching SRT unless the user opts out. Validate monotonic non-overlapping cues, at most two lines, mobile-safe margins, and cue end at or before the master.

Use the approved brand brief for opening-overlay duration; without one, use four seconds as a starting point, or a clean opening when the source calls for it. An overlay cannot repair a weak spoken hook. Reinforce the spoken promise without strengthening the claim. Check collisions with captions, lower thirds, chapter cards, and qualifiers. Claim qualifiers take priority.