⚙️ The skill under the hood

The engine detects speech segments using audio energy (silencedetect) and removes silences, pauses, repetitions, and mistakes, keeping only the final take of each sentence.

It’s the foundation: a predictable cut that’s always the same for the same input—no timing guesswork.

silencedetect, voice island, last take, determinism.

The engine looks at the frames (via /watch), understands the content, and suggests the opening, B-roll type, and pacing—putting together an editing plan (beat sheet).

It’s what makes the treatment adapt to THE video, instead of always applying the same template.

/watch, treatment, beat sheet, opening.

Hyperframes builds the animations, and Firecrawl captures real websites/repos. The cold open is mandatory, and no segment goes more than 4s without a visual hit.

These are the phases that give it a "pro level": rhythm, credibility from real footage, and a hook right at the start.

Hyperframes, Firecrawl, cold open, 4-second rule.

Subtle sound effects (whoosh, ching) play under the voice—no music by default. The file is rendered, and an automatic QC check looks for gaps and overlaps, regenerating anything that fails.

Well-placed SFX "creates" the cut; QC prevents you from delivering a render with a sync gap.

SFX, render, QC, regeneration.

A clean subagent watches the render (via /watch) and gives a PASS or FAIL verdict. If it fails, the engine fixes it and repeats the cycle until it passes.

It’s the last gate before delivery—nothing goes out without independent approval.

Reviewer, subagent, PASS/FAIL, loop.

The reviewer is a subagent that didn’t see how the video was assembled. It judges only the result, without the bias of the person who built it.

The person who assembles it tends to "not see" their own mistakes; a fresh pair of eyes catches what the main agent missed.

Clean context, author bias, independent verification.

Three ideas that tie the track together: the hard gate (blocking verification), the beat sheet (rhythm plan), and the quality loop (repeat until it passes).

With this vocabulary, you understand why the engine is reliable, not just “smart.”

Hard gate, beat sheet, quality loop.

The cut uses the exact silence boundaries in the audio, not Whisper’s timestamps (which have ±0.2–0.3s of jitter). The islands.py script finds the speech islands.

Cutting with Whisper leaves bits of silence and half-words; cutting based on energy is precise.

silencedetect, jitter, island, islands.py.

islands.py proposes what to keep (KEEP/DROP), cut.py assembles the cut, and verify-cut.py checks that it’s clean — exiting with exit 0 when it passes.

It’s a chain with a check at the end: you don’t animate anything before the cut passes.

KEEP/DROP, exit 0, hard gate.

A static lint for the Hyperframes timeline: if there are more than 4s without a visual beat, it reports an error and prevents rendering.

The 4s rule stops being an intention and becomes an automatic pacing check.

Lint, visual beat, --max-gap, pacing gate.

Generates 2–3-word caption “beats,” synced to the voice, with a keyword (which you pass in --keywords) highlighted in your accent color.

Short captions with their own visual emphasis avoid the generic “AI look” of plain captions.

Caption beat, keyword, safe area.

make-sfx.sh synthesizes the sound effects palette (whoosh, boom, ding…) without downloading anything; mix-sfx.py overlays each effect at the right time, under the voice, with a limiter.

Synthesized SFX are lightweight, royalty-free, and always the same — they fit the deterministic pipeline.

SFX palette, events, ducking, limiter.

Animations don’t use Math.random() or Date.now(); no infinite repeats (repetitions are finite and calculated); timelines are registered in a paused state.

Only this way will the same project render the same frame every time — a requirement for reliable lint and QC.

Determinism, finite repeat, paused, reproducibility.