produce-video

Produce Video

Take a recording the user provides and turn it into a finished video. The recording is the spine and drives the timeline; this skill edits (silence + chosen mistakes/pacing) and augments (overlays), then renders and generates launch content. It does NOT author a synthetic video — for PR-driven synthetic promos use hyperframes-video instead.

Quality principle (non-negotiable): every footage step preserves the original quality. Never let a tool re-encode the footage at low quality. De-silence and cuts run at near-lossless CRF 12 from the original (auto-editor is used only to DETECT silence, never to encode — its default crushes footage to ~0.6 Mbps), and the render runs at CRF 12. CRF gives small files for static screencasts; that is full quality, not loss.

Relation to other skills (delegate, don't re-derive)

• /transcribe-video — all transcription (returns word-level transcript.words.json). Invoke it; do not transcribe by hand.
• hyperframes — composition authoring rules (Visual Identity Gate, Layout Before Animation, timeline contract, video/overlay layering). Invoke before writing/mounting any composition HTML.
• hyperframes-cli — every npx hyperframes command (init, lint, inspect, preview, render).
• gsap — GSAP timeline/easing patterns for overlays.
• youtube-copy — YouTube title/description/tags/chapters (P6).
• blog-post — the blog post (P6).
• social-copy — X/Bluesky/LinkedIn/Reddit copy (P6).

Bundled assets

• scripts/desilence.mjs — auto-editor silence removal + before/after duration.
• scripts/extract-frame.mjs — single-frame grab (framing detection + on-screen highlight).
• scripts/bake-cuts.mjs — ffmpeg trim/concat cutter (dense keyframes) used in P2.
• templates/overlays/*.html — 9 overlay sub-composition templates (filled per instance; sub-comps are standalone HTML docs in the installed CLI — read each file's header comment).
• templates/project/{DESIGN.md,styles.css} — Visual Identity Gate + brand var(--brand-*) tokens the overlays consume.
• references/ — mistake-detection.md, overlay-triggers.md, framing-safe-zones.md, frame-analysis.md.
• P6 content generation is delegated to the youtube-copy, blog-post, and social-copy skills (no vendored LLM script).

Gates (non-negotiable)

1. P2 mistake review is VISUAL, never terminal — whenever ≥1 flag exists, the agent ALWAYS builds the hyperframes marker preview (markers on the video at each flagged span) and launches it, so the user decides by watching. Do NOT present the flags as a terminal table/question for the user to decide from, and do NOT make the preview optional or ask "want me to show you?" — just build it and open it. A short terminal summary of the flags is fine, but the decision surface is the preview. Never auto-apply a cut/ramp.
2. P4 overlay plan approval — present the timestamped overlay table; no build until approved.
3. Preview is always agent-launched — for both the P2 review and the P5 review, the agent starts npx hyperframes preview and opens the browser tab itself. NEVER ask the user to run preview or open a URL, and NEVER substitute a terminal-only summary for the live preview.
4. Explicit-render HARD-GATE — render runs ONLY on an explicit user render command ("render", "ship it"). After a render, edit requests return to the preview loop and do NOT re-render until the user explicitly says so again. Every render needs its own fresh command. This overrides any "use sane defaults / non-interactive / just ship it" phrasing.

"Use sane defaults" / "don't ask questions" skips the configuration questions (P0) and the iteration prompting — it NEVER skips the approval gates, the agent-launched preview, or the explicit-render gate.

Process

Preflight — check dependencies & skills (run FIRST, before P0)

Verify the environment before doing any work; warn the user about anything missing and how to fix it.

• auto-editor (required for P1 de-silence): run auto-editor --version. If it is not installed, STOP and tell the user: "auto-editor is not installed - install it with pipx install auto-editor (or pip install auto-editor), then re-run." Do not attempt P1 without it.
• ffmpeg / ffprobe (REQUIRED — P2 cuts, P4 frame analysis, P5 render, audio): run ffmpeg -version. If missing, STOP and tell the user to install ffmpeg and add it to PATH. The pipeline cannot run without it.
• Docker (for /transcribe-video): docker info. Warn if the daemon is down (start Docker Desktop).
• Node >= 22: node --version. Warn if older.
• Required skills — confirm these are available (try to invoke / check the skill list); warn for any missing and which phase it blocks:
  • transcribe-video (P2/P3 transcription), hyperframes + hyperframes-cli + gsap (P2/P4/P5 composition & render), youtube-copy + social-copy (P6). blog-post only if the user opts into a blog (P6, optional).
• Summarize the preflight result in one short block (what's present, what's missing). Only proceed past a missing required dependency/skill if the user explicitly says to continue without it (and note the degraded result).

P0 — Input & configuration

• Input: path to the raw recording (required).
• Optional: a repo/files for real code grounding; visual style/brand (else run the hyperframes Visual Identity Gate — detect from the repo, else ask 3 style questions; write templates/project/DESIGN.md + brand values into styles.css); overlay density (sparse / balanced / rich).
• Single output folder (<OUT>): everything this skill produces lives under ONE folder, default <recording-dir>/<video-slug>-produced/ (override with a user-supplied path). Create it now. Final structure:

  <OUT>/
    final.mp4                 # rendered video (P5)
    hyperframes/              # the HyperFrames project: index.html, styles.css, DESIGN.md, compositions/, renders/ (P4–P5)
    transcript.txt            # final edited transcript (P3 / P6)
    transcript.srt
    transcript.words.json
    youtube.md                # P6
    socials.md                # P6
    blog.md                   # P6
    .work/                    # intermediates: 01_desilenced.mp4, 02_edit_N.mp4, frames/ (prunable in P7)
  

  All subsequent phases write inside <OUT> — never scatter outputs elsewhere. <work> below = <OUT>/.work.

P1 — De-silence (lossless)

node scripts/desilence.mjs <input> --out <work>/01_desilenced.mp4 --margin 0.3s

desilence.mjs uses auto-editor ONLY to detect the silent ranges (--export v3), then cuts the ORIGINAL with ffmpeg at near-lossless CRF 12, keeping the original audio. Do NOT let auto-editor encode the output — its default re-encode crushes the footage to ~0.6 Mbps and permanently blurs it (faces especially); no downstream bitrate can recover that. Report before/after duration and seconds removed. This becomes the working copy.

P2 — Mistake & pacing review loop (approval gate)

1. Transcribe the working copy: invoke /transcribe-video on <work>/01_desilenced.mp4 with --out <work> → working transcript (intermediate; superseded by the P3 final transcript).
2. Detect flags per references/mistake-detection.md (cuts + speed-ramps), snapped to word boundaries.
3. Always (whenever ≥1 flag exists) build the visual review. Scaffold a lightweight hyperframes project with the de-silenced video as the base track; place visually-distinct temporary markers ON THE VIDEO at each flagged span (cut = red strike, ramp = amber fast-forward, each labeled with the excerpt + its index). Agent starts npx hyperframes preview and opens the browser itself (per references/framing-safe-zones.md). This is the decision surface — never replace it with a terminal table. A one-line terminal note ("3 flags marked in the preview — review and tell me which to cut/ramp/keep") is the most the terminal should carry.
4. The user watches the marked preview and picks per span: cut / speed-ramp (factor, default ~1.5–2×) / leave / add their own ranges.
5. Bake selections → <work>/02_edit_N.mp4, then re-preview:
   • Cuts: use node scripts/bake-cuts.mjs <input> --out <work>/02_edit_N.mp4 --cut s,e --cut s,e … (ffmpeg trim/concat at near-lossless CRF 12 by default — trims video+audio identically in one pass). Do NOT use auto-editor --cut-out for the mistake cuts — its variadic arg mis-assigns the last range as a positional input, it re-encodes lossily, and a select-filter approach desyncs audio.
   • Speed-ramps: auto-editor --set-speed-for-range speed,s,e (pitch preserved); pass ONE range per invocation, or place --set-speed-for-range last on the command line, to dodge the same variadic-arg pitfall.
   • All cut/ramp timestamps are on the de-silenced timeline (the transcript the flags came from). Apply cuts and ramps in the same round against that timeline.
6. Loop until the user says "done". Drift guard: after ~10 rounds offer to reset to an earlier intermediate. Final artifact: <work>/final_cut.mp4.

Cuts and ramps are baked into the footage HERE (before P3) so the final transcript's word timestamps match the edited timeline.

P3 — Final transcription

Invoke /transcribe-video on <work>/final_cut.mp4 with --out <OUT> so transcript.txt / transcript.srt / transcript.words.json land at the top of the output folder. This word-timestamped transcript is the sync source for all overlays (P4) and the input to content generation (P6).

P4 — Overlay design & build (approval gate)

1. Framing detection (references/framing-safe-zones.md): sample frames with extract-frame.mjs, classify phases, confirm phases + safe zones with the user.
2. Overlay-trigger detection (references/overlay-triggers.md): scan the word-timestamped transcript → overlay plan rows {start, est_duration, type, content, safe_zone, why}. For on-screen code, use references/frame-analysis.md (extract frame → locate → normalized box → stability guard). For code cards, repo-if-available-else-synthesize (flag synthesized).
3. Self-improvement pass (≥2 iterations): honor density, enforce gaps + dwell minimums, validate every overlay against the active phase's safe zone.
4. Approval gate: present the timestamped table; user edits/removes/retimes/adds. Synthesized code shown for verification. No build until approved.
5. Build (invoke hyperframes + gsap first):
   • npx hyperframes init <OUT>/hyperframes --video <OUT>/.work/final_cut.mp4 (the project lives at <OUT>/hyperframes; base: video track 0 muted playsinline; audio on a separate track at volume 1 so the voice plays). Read the generated index.html to learn this CLI version's sub-composition include syntax.
   • Copy templates/project/styles.css (with the confirmed brand values) into <OUT>/hyperframes.
   • For each approved row: instantiate the matching templates/overlays/*.html, give it a unique composition id + data-composition-id, fill its {{tokens}}, set data-start = the transcript timestamp and data-duration = the dwell, mount it above the video. Punch-in zooms scale the video's WRAPPER div (never the <video>).
   • Output dimensions inherit the source recording.
   • Readability (non-negotiable): all overlay text must be clearly legible over ANY video background. Every text group sits on a near-opaque dark panel (rgba(8,10,14,0.92), rounded, subtle border/shadow); titles/body are white (var(--brand-text)), secondary/kicker/meta are light slate (#CBD5E1). NEVER use var(--brand-accent)/var(--brand-primary) for body/subtitle/kicker text — accent is for bars, borders, marker sweeps, icons, and big display words only (and only when they clear ~4.5:1 on the panel). When unsure, make text white. npx hyperframes validate (WCAG contrast) must pass with zero contrast warnings.

P5 — Preview & render (mandatory agent-launched preview + explicit-render gate)

• npx hyperframes lint + npx hyperframes inspect (overflow + safe zones) + npx hyperframes validate (contrast). Fix until clean.
• Agent starts npx hyperframes preview and opens the browser. Freeform iteration loop on overlays.
• Pre-render audits: every overlay in its phase's safe zone; dwell minimums; no two overlays fighting the same region simultaneously; lint/inspect/validate clean.
• On the user's explicit render command only: render at near-lossless CRF 12. The default standard quality encodes at ~1.4 Mbps and visibly crushes footage; always pass a low CRF:

  npx hyperframes render --output <OUT>/final.mp4 --crf 12
  

  (run from <OUT>/hyperframes. CRF 12 is near-lossless; do NOT use --video-bitrate expecting a big file — for static screencasts CRF produces a small file at full quality, and that is correct, not low quality.) Final mp4 keeps original audio. After render, further edits return to the loop and require a fresh explicit render command.
• Quality note: the Chrome capture is a lossless 1:1 screenshot (verified: render input == output for both text and faces), so a low-CRF render preserves the footage it is given. The ONLY thing that destroys quality is a lossy upstream step — which is why P1 de-silence and P2 cuts both run at near-lossless CRF 12 from the original. Do not add extra re-encode passes. For absolute zero-generation footage, an advanced path is to render overlays-only to a transparent --format webm/mov and ffmpeg-composite them over the cut footage (punch-in zoom then applied via ffmpeg) — rarely needed.

P6 — Content generation (delegated to specialized skills)

Generate the launch content by invoking dedicated skills, each on the final edited transcript (P3 — pass the timestamped transcript.srt so YouTube chapters get accurate times), writing each output into <OUT>:

• youtube-copy → <OUT>/youtube.md (click-worthy title, above-the-fold description, tags, and timestamped chapters from the SRT). Default on.
• social-copy → <OUT>/socials.md (X + thread, Bluesky, LinkedIn, Reddit). Default on.
• blog-post → <OUT>/blog.md (full blog post from the transcript). Optional — ask the user "Want a blog post too?" and only run it if they say yes (default: skip). A blog is a bigger artifact many videos don't need.

Invoke each via the Skill tool with the transcript as the source; they own their own quality rules. Do NOT use a monolithic LLM script for this. If a skill is unavailable, note it and continue with the others (the rendered video is already saved).

Writing rule for ALL generated text (titles, descriptions, chapters, blog, socials): never use em-dashes (— / –), use a hyphen -.

P7 — Cleanup & handoff

• Confirm <OUT> contains the consolidated result: final.mp4, hyperframes/, transcript.*, youtube.md, socials.md, blog.md.
• Offer to prune <OUT>/.work/ (intermediates: de-silenced/cut files, extracted frames). The user owns disposition.
• Open <OUT> in the file explorer for the user and report the single folder path.

Requirements

Docker (for /transcribe-video), auto-editor + ffmpeg, Node ≥22, a HyperFrames-capable environment, and an LLM provider for P6.