Longform Video

longform-video — the mixed-format longform line

Proven 2026-08-29 (v5, 4:50, exit=0). Every rule below was paid for that night or earlier. Read /two-brents-brand FIRST for voice/story/compliance. This skill is the PRODUCTION layer.

The format (locked by Brent 08-29)

1. Picture = three interleaved sources:
   • Brent's REAL footage (hook, teaching-on-camera, ground/CTA, demo) — never re-cut mid-take without transcript timings.
   • Faceless HyperFrames beats for AI-Brent teaching (via /faceless-explainer) — Broadside-class ink/brass kinetic type. Brent 08-29: "I love what you did with the graphics."
   • Blue-shirt clone appearance (HeyGen registry ai_brent 44f56d7e) in 1–2 banter beats so viewers SEE who's talking — "nobody knows what the hell is going on, it's just my voice." Not every beat. Existing rendered modules first; NEW HeyGen render = ping Brent the dollar amount BEFORE spending, free test:true + his eyeball before credits (registry law).
2. Voice = Brent's Voicebox clone (profile 18836009-0527-4b9f-9837-3d58c2a0dc76, /brag-voice) for all faceless narration. Clone modules carry their own HeyGen ElevenLabs voice — acceptable per Brent 08-29, keep appearances short.
3. Cards — LOCKED, from Youtube-system (~/Desktop/YOUTUBE/hyperframes-cards/, repo bbrysonelite-max/Youtube-system; read CARDS-LOCKED.md before trusting any render):
   • FRONT intro/frontcard-uniform.mp4 (5s) — the follow/"keep the lights on" ask. Place after the hook, or at the first clean cut near the 90-second mark.
   • BACK newest endcard/endcard-v*.mp4 (14s) — SUBSCRIBE + type-CLAW + brentbryson.ai SPELLED OUT. Never ship a homemade "link below" card: every CTA card must show a typed URL or word people can spell.
   • tap-card/tap.mp4 — CLAW conversion, optional mid-video.
   • Stale-render trap: after editing a card source, re-render AND look at a frame.
4. Compliance — python3 ~/Desktop/YOUTUBE/scripts/compliance_check.py <every text> before render. No company name near any dollar figure, ever, on-canvas included ($25M stands alone). Public fence words apply.

Pipeline (each gate has a receipt or you stop)

1. Script gate — Brent-approved script, verbatim (VO_MODE verbatim). No script, no build.
2. VO — generate/reuse Voicebox wavs per AI beat. ffprobe every wav; faceless frame duration = wav + 0.4–0.6s pad.
3. Faceless project — /faceless-explainer INSIDE the video's system folder (videos/<project>/), one folder per system. Mark SILENT (music: none, no SCRIPT.md) — pipeline TTS default is a stranger's voice; never let it generate narration. One Opus worker per frame, packets via frame-packets.mjs. Fonts: pin identical @font-face gstatic woff2 blocks in EVERY frame (workers diverge — normalize before lint) + vendor woff2 into assets/fonts/.
4. Checks to zero — npx hyperframes lint + check = 0 errors. Known traps: letterSpacing/layout-prop tweens (stutter — transforms only), multi-writer GSAP tweens on one property (seek renders camera frozen — ONE driver tween per property), #root .clip specificity overriding positioned .clip elements, exits at clip boundaries need tl.set hard kill. Overlap errors: mark data-layout-allow-overlap ONLY after looking at the frame.
5. Render — npx hyperframes render --quality high (local, $0). Snapshot contact sheet, LOOK at it.
6. Assembly — script file, never inline (rtk mangles ffmpeg vars; bash $((...)) treats "09" as bad octal — no numeric-string array keys). Re-encode EVERYTHING to one spec before concat: 1920x1080, 30fps, libx264 crf19, aac 192k 48k stereo; upscale 720p cards with scale+pad; add anullsrc silence to soundless cards; mux VO with -map 0:v -map 1:a -af apad -shortest. Concat with -c copy over the list file.
7. QC (real eyes) — pull frames at faceless, real-footage, clone, and card offsets; volumedetect at VO and footage spots (expect ≈ -20 to -30dB mean, same ballpark both). Compute offsets from per-segment ffprobe durations, never from memory.
8. Chapters + metadata — re-measure chapter marks from actual segment durations every re-cut; description carries the C1 canonical block.
9. Brent gate — he watches the file. NEVER upload; upload is his. Rung 4 until his eyes = rung 5.

Hard laws inherited

• No music bed decision inside faceless renders; bed (if any) at final assembly.
• Shorts = cut from THIS longform only.
• ffmpeg build here has NO drawtext; cards via PIL or HyperFrames.
• Whisper lane = brew whisper-cli + ggml-base.en.bin (Voicebox /transcribe is broken).
• Deliver ONE file per iteration (v1, v2 …), keep prior versions, never overwrite his kept cut.