Leveling Up Content Engine

Leveling Up 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.

Updated editorial and release workflow

Read the current workflow updates before selection, editing, packaging, review, or release planning. These rules supplement the bundled format, opportunity, quality, and delivery contracts.

Use the source's actual show and brand. A derivative retains its parent show even if only one speaker appears. Obtain the authorized brand brief and assets; this package includes no private channel access or internal recording history.

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.

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.