Imported from novoads/agent-skills (shared/skills/broll-overlay/SKILL.md). Install upstream with npx skills add novoads/agent-skills --skill broll-overlay. Copyright stays with the author.

B-roll overlay

Lays silent cutaway clips over a finished base video: the base audio runs untouched underneath, the picture cuts away and returns, and the final duration equals the base duration exactly. That is the whole contract — overlay, never concatenation. A previous run concatenated: 24s for a 15s base, voice dead 12s.

Assembly mechanics live entirely in scripts/broll_overlay.py; this file is judgment only — when to enter, where the windows go, what the user sees first.

Entry condition

Two things must already exist. This skill never generates the base — if there isn't one yet, say so and stop.

1. A finished, approved base video with exactly one audio stream. Approved means the user has watched it and signed off — b-roll is a polish pass over a locked cut. The script stream-copies one voice track under the picture, so a two-stream base is rejected: ffmpeg -i base -map 0:v -map 0:a:0 -c copy.

2. Its transcript, with timings. One call to POST /v1/transcripts with the base's jobId (or its assetId if it was uploaded). It returns text, word-level words[] and segments[] — timings in SECONDS, which is what the EDL below takes — plus an srt you can ignore here. Price it with the transcript arm of POST /v1/estimates and quote that number, never one from this file. The meter is per minute of source, rounded up from a one-minute minimum, and a repeat of the same source is free, so asking twice costs once. No local install.

   (The captions endpoint still cannot supply this — it burns subtitles into a new MP4 and returns no text or timings, verified 2026-08-04 and re-checked when transcripts shipped. That finding is correct and is kept here so nobody re-discovers it and files a bug against captions.)

   Offline fallback: a local whisper pass (whisper-cli or openai-whisper) still works and is the path when there is no API key or no network. Note whisper reports timings in milliseconds where the API reports seconds, and whisper-cli with no model downloaded returns an EMPTY transcript rather than an error — which reads exactly like a bad render. See README for the model download.

   Placement is read from the transcript, not eyeballed.

REST key required. A Novoads MCP connector is not a substitute. If NOVOADS_API_KEY is missing or still the placeholder, stop before any generation work and tell the user: "Before continuing, create an API key at https://novoads.ai/dashboard/settings?tab=api and paste it into .env." That holds even when mcp__novoads__* tools are connected and authenticated in the session. Never call mcp__novoads__* tools from this repo's workflows: they are a different surface with different behavior, including the units they quote costs in. Repo installs verify with ./scripts/check-novoads-env.sh; a solo install checks NOVOADS_API_KEY in the environment.

Pack version. Every /v1 response carries X-Novoads-Pack-Version; mention a newer pack at https://github.com/novoads/agent-skills only when that header names a version NEWER than this file's metadata.packVersion — equal or older is nothing to say, and it is never a reason to stop.

Workflow

1. Read the transcript. Note what is said and when — the words drive every decision below.
2. Generate the b-roll clips, only after base approval. One clip per moment you intend to cover, each rendered silent (audioEnabled: false). Overlay audio is ignored by the script regardless, and generating before approval spends credits on a cut that may still change. Price the set live first — POST /v1/estimates this session, the total shown against the user's balance, an explicit yes. Four cutaways is four charges, and a remembered price is not a price. Cast the base actor in every cutaway that shows a person: pass the same referenceAssetIds the base used — Seedance re-casts every render, and a stranger's hands or skin breaks the ad (observed live 2026-08-04: Black base actor, five white-cast cutaways). With no identity ref, keep people out of frame entirely — objects and environments, not "face out of frame".
3. Propose an EDL. JSON with base, output, and overlays (file, start, end, covers). covers is mandatory and quotes the spoken line the window illustrates — a window you can't quote is one you're guessing at.
4. Show the EDL to the user and wait for approval. Rendering before the user has seen the plan is precisely the failure this step prevents.
5. Render: python3 scripts/broll_overlay.py edl.json
6. Report the script's verification output verbatim — the measured durations, the audio check and the window count, not a paraphrase. Then offer 2–3 cut variations (different windows, fewer cutaways, a tighter opener) to react to.

Placement judgment

Default cadence, measured 2026-08-04 frame by frame from the reference edit this pack reproduces — ~15.5s, 10–11 shots, a cut every ~1.4s:

• Alternate A-B-A-B: the face returns between every cutaway. Never two overlay windows back to back; the talking head is the spine.
• 4–6 windows per 15s, ~1–1.5s each, ~40–45% coverage. Short and frequent. This retires our older "fewer, longer windows" rule — taste, reversed by measurement; our own two renders ran 2 windows of 2.5s/2.0s, ~30% coverage.
• The b-roll may travel; the base never does. The reference visits 4 distinct b-roll settings across 2–3 rooms; the talking head never moves.
• Never cover the opening hook beat. The first beat is a face making a claim; cutting away there spends the retention the hook just bought.
• End on the person, not on b-roll. The closing beat is the verdict, and it only lands if the viewer is looking at whoever delivers it.
• Put product shots where the product is being spoken about. B-roll that illustrates the current sentence reads as evidence; anything else is filler.

Default shot plan. The reference's five cutaways tell one arc: problem → stress → product macro → dose/usage → relief — 2 product shots to 3 emotional beats. Adapt the imagery to the script, keep the ratio.

Evidence: one measured edit (n=1) — a default, not a law; depart deliberately and say why (evals.md OV3/OV6). Every run prints its own cadence against this envelope; --stats prints it without rendering.

Hard rule: assembly always goes through the script

Never hand-write overlay or concat ffmpeg for this task. broll_overlay.py owns validation, rendering and verification, and it fails loudly on exactly what fails silently by hand: duration drift, re-encoded audio, overlapping or zero-frame windows, a clip shorter than its window, a rotated phone base composited at the wrong geometry, a window that quietly composited nothing. Improvised ffmpeg is how the founding failure happened.

It needs ffmpeg 7.1 or newer. The script scales each overlay against the base branch with scale=w=rw:h=rh, and those reference constants arrived in 7.1. On an older build — Ubuntu 24.04's apt ships 6.1 — the graph does not parse and the error names rw, not a version. Check with ffmpeg -version before blaming the EDL; on Debian/Ubuntu the fix is a static build, not apt install ffmpeg.

Escape hatch: if a request genuinely doesn't fit the EDL model — a speed ramp, a picture-in-picture inset, audio that actually needs editing — say so and ask the user how to proceed. Do not improvise around the script.

Other modes

• python3 scripts/broll_overlay.py edl.json --dry-run — validate the EDL and print the plan, rendering nothing. Cheap way to check window geometry.
• python3 scripts/broll_overlay.py edl.json --stats — window count, lengths, coverage, base-return gaps, each marked against the envelope. Never an error.
• python3 scripts/broll_overlay.py --verify-only FINAL --base BASE [--edl edl.json] — re-check any output against its base, including one from an old session. Pass --edl whenever you have it: without it the check is duration + audio only, which a plain copy of the base also passes; with it each window's midpoint frame is compared against the base, so a window that composited nothing fails. A run without --edl says so out loud.

Exit codes: 2 validation, 3 render, 4 verification. A nonzero exit is a real failure — surface it, don't retry blindly. Renders go to a hidden temp file beside the output, renamed only after verification passes: a failed run never leaves a half-written or unverified file at the output path.

Evals

evals.md defines the six scenarios this skill is held to (OV1–OV6). scripts/test_broll_overlay.py implements the mechanical ones against synthetic fixtures — run it after touching the script.