anton-abyzov/vskill

/hyperframes-best-practices

Use this skill any time you are turning a raw screencast, talking-head recording, or OBS export into a polished HyperFrames product video. It bundles two layers of knowledge:

1. HyperFrames fundamentals — the upstream rule files maintained by HeyGen at github.com/heygen-com/hyperframes (composition authoring, GSAP timelines, data-* attributes, the dev-loop CLI). Install via npx vskill i heygen-com/hyperframes.
2. Production-craft rules — lessons captured from a long iteration cycle building a real SpecWeave product video from a single 8:47 OBS recording. These are the rules that, when followed, produce a video the user can actually ship — and, when broken, leave you with beeps, mid-word clips, blank takeover scenes, and overlays that block the speaker's face cam.

Read both layers before authoring.

---

Scope and applicability

The Layer 2 craft rules below were derived from one specific project: turning an 8:47 SpecWeave intro recording (Anton Abyzov, OBS capture, talking-head + screen-share, May 2026) into a 6:00 production-ready promo video. They reflect the look-and-feel of that product — Studio's tinted-card aesthetic, SpecWeave's purple/green palette, an iOS-card overlay style, and the technical realities of OBS recordings (start/stop ticks, recording-resume transients, duplicate takes from re-recording the same line three times).

The exact constants are project-specific — overlay positions for that face-cam location, the Boris Cherny / Anthropic correction text, the 600+/500+/3K+ stats — translate, do not copy.

Use this layer as a reference for the underlying technique: the transcript-driven smart-cut algorithm, the per-clip fade-in for transient suppression, the con​cat-filter render pipeline that avoids con​cat-demuxer clicks, the surgical-mute beep scrubber, the iOS-card overlay pattern, the source-time → cut-time mapping, the Studio waveform-cache invalidation gotcha — those generalize.

---

When to use

Trigger this skill when ANY of these apply:

• A user has a raw screencast or OBS recording and wants a tightened, captioned, branded video.
• The source has dead air, paused silences, or duplicate "let me try that again" takes.
• The source has clicks, beeps, or transient artifacts at recording boundaries.
• The user wants graphic overlays (callouts, stat cards, brand reveals, tool logos) that don't cover the speaker's face cam.
• The user said the words "make this production-ready" or "ready to ship".
• The user mentioned wanting "Apple-keynote style" or "iOS hover cards" or "frosted glass".

If ANY of those signals are present, follow this workflow. If the user just wants a quick title card on a clean source, the upstream hyperframes skill alone is enough.

---

The pipeline (10 phases)

┌──────────────────────────────────────────────────────────────────┐
│  Phase 0  Inventory tools                                         │
│  Phase 1  Probe the source (duration, codec, audio sample rate)   │
│  Phase 2  Transcribe with Whisper (word-level timestamps)         │
│  Phase 3  Detect silences + scene changes                         │
│  Phase 4  Build the smart cut list (with word-boundary buffers)   │
│  Phase 5  Scrub transient beeps (surgical mute + adeclick)        │
│  Phase 6  Render the master via con​cat filter (single pass)       │
│  Phase 7  Generate hero imagery (Gemini 3 Pro Image)              │
│  Phase 8  Author the HyperFrames composition (iOS cards)          │
│  Phase 9  Lint + verify in Studio                                 │
└──────────────────────────────────────────────────────────────────┘

Run the phases sequentially. Iterate inside a phase, not across phases — a bad cut list breaks every overlay timing downstream, and you'll spend an hour re-resolving anchors. Get cut list right first, then render once, then build overlays against the final timeline.

---

Phase 0 — Inventory tools

Before starting, confirm these are available. The skill assumes macOS / Linux; on Windows use WSL.

Tool	Used for	Install
ffmpeg (>= 6.0)	All video/audio processing	brew install ffmpeg
ffprobe	Stream inspection	bundled with ffmpeg
whisper (OpenAI Python pkg)	Transcription with word timestamps	pip install openai-whisper
python3	Cut list / beep scrubber scripts	system
npx hyperframes@latest	Composition CLI (preview, render, lint)	npx hyperframes init <project>
Gemini 3 Pro Image API	Hero images (optional)	GEMINI_API_KEY env var
vskill (>= 1.0)	Skill installation	npm i -g vskill

If Whisper isn't installed, pip install openai-whisper brings it in. Do not use large-v3 or large-v3-turbo on CPU — they take 10–15 minutes for an 8-minute audio. Use base.en for English tech monologue: ~30s, plenty accurate for word boundaries.

If a user's machine has auto-editor installed, ignore it — it does silence cuts but not beep scrubbing or word-boundary buffering, and it generates choppy 50+ segment cuts. We do better by hand.

---

Phase 1 — Probe the source

Before doing anything else, learn the file's shape:

ffprobe -v error -show_entries stream=index,codec_type,codec_name,width,height,r_frame_rate,duration,sample_rate,channels \
  -show_entries format=duration,size,bit_rate -of json "$SRC" | tee source_meta.json

Things you must know:

• Total duration — drives all cut math.
• Frame rate — OBS often records at 60 fps; HyperFrames Studio plays at 30 fps internally for scrubbing. Multiply seconds × 30 to get studio frame numbers, not seconds × source-fps. This trips people up.
• Audio codec & sample rate — OBS commonly records pcm_s24le at 48 kHz. Re-encode to AAC 192k for delivery; keep 48 kHz to match all downstream tools.
• Container — .mov from DaVinci, .mkv from OBS — both fine, ffmpeg reads either.

Save the probe to disk so later phases can re-read it without re-probing.

---

Phase 2 — Transcribe with Whisper

Word-level timestamps are the foundation. Without them you can't:

• Snap cut boundaries to word edges (avoiding mid-word clips).
• Detect duplicate takes (same line said 2-3 times).
• Place overlay anchors against speech moments instead of cut times that drift.

Extract audio first so Whisper gets a clean mono 16 kHz WAV (saves 80% of CPU vs full-quality):

mkdir -p .work
ffmpeg -hide_banner -y -i "$SRC" -vn -ac 1 -ar 16000 -c:a pcm_s16le .work/source.wav

Run Whisper with word timestamps and JSON output:

whisper .work/source.wav \
  --model base.en \
  --language en \
  --word_timestamps True \
  --output_dir .work \
  --output_format json \
  --verbose False

⚠️ Whisper CLI gotcha: passing --output_format json --output_format srt keeps only the last flag in older versions. If you need multiple, use --output_format all (writes JSON + SRT + VTT + TSV + TXT).

Validate the output: .work/source.json should contain segments[].words[] with start, end, word. If words is empty, you forgot --word_timestamps True.

Read the full transcript at this stage. Don't skip it. You'll spot:

• Brand-name miss-hearings. Whisper hears "Claude Code at Anthropic" as "Cloud Code at Infropic" 90% of the time. "SpecWeave" comes out as "spec with" or "specs". These need overlay corrections, not re-recording.
• Duplicate takes. Look for the same sentence repeated verbatim. In our project, "And those systems cannot produce skills as you work" appeared three times at 01:14, 01:30, 01:47. Drop two of them.
• Filler restarts. Single-word fragments like "And…", "And…" preceding the actual sentence. The cut script catches these as "ghost" segments — we drop them in phase 4.

---

Phase 3 — Detect silences and scene changes

Two parallel detections feed the cut algorithm:

3a. Silence detection

ffmpeg -hide_banner -nostats -i "$SRC" -af "silencedetect=noise=-30dB:duration=1.2" -f null - 2>&1 \
  | grep -E "silence_(start|end)"

-30 dB and 1.2s are the sweet spot. Looser (-25 dB, 0.8s) catches breath gaps you don't want to cut. Tighter (-35 dB, 2s) misses real dead air.

Parse start/end pairs into [(s, e, dur), ...] and write to silences.json. See scripts/build_cuts.py.

3b. Scene-change detection

ffmpeg -hide_banner -nostats -i "$SRC" -filter:v "select='gt(scene,0.05)',showinfo" -f null - 2>&1 \
  | grep "showinfo" | grep -oE "pts_time:[0-9.]+" | sed 's/pts_time://' > scenes.txt

scene>0.05 is sensitive enough to catch the speaker bringing up a new browser tab, switching apps, scrolling significantly. Don't go below 0.03 — you'll get false positives from cursor blinks.

These two together let you make the smart cut decision: silence + no scene change = pure dead air, cut it. Silence + scene change = the speaker is showing something on screen, keep it (the audience is supposed to look at the screen, not hear narration).

---

Phase 4 — Build the smart cut list

The cut algorithm has 5 hard rules earned the painful way:

Rule 1 — Pure dead air gets cut, but not all of it

For each silent segment > MIN_CUT_LEN (1.5s) with no scene change inside: cut everything except a BREATH of 0.30s. The breath preserves natural rhythm — eliminating 100% of silence makes speech sound machine-edited.

if not in_scenes and (cut_end - cut_start) > MIN_CUT_LEN:
    cuts.append((cut_start + 0.30, cut_end, f"dead air {dur:.1f}s"))

Rule 2 — Silences with screen activity stay (capped at 3.5s)

If the silence has scene changes, the speaker is letting the audience read. Keep the window from first_scene - 0.4s to last_scene + 0.8s, but never longer than 3.5s — staring at one frame past 3.5s loses people.

Rule 3 — Snap to word boundaries

After computing the rough keep ranges, snap each boundary to the nearest word edge from Whisper, within a 0.4s tolerance:

def snap(t, words):
    best, bd = t, 0.4
    for s, e, _ in words:
        for cand in (s, e):
            if abs(cand - t) < bd:
                best, bd = cand, abs(cand - t)
    return best

This is what stops cuts from clipping mid-word. Without it, every other cut clips a trailing 's' or breath consonant.

Rule 4 — Add word-boundary buffers (this is the one users notice)

Even with word-edge snapping, trailing consonants extend slightly past Whisper's reported word-end timestamp. Whisper times the vowel onset to vowel offset, not the consonant tail. So a literal cut at word.end clips the trailing 's', breath, or 'k'.

The fix: extend each keep range:

• End: + 0.30s past the last word's end, capped at half the gap to the next keep.
• Start: - 0.10s before the first word's start, capped at half the gap to the previous keep.

END_BUF, START_BUF = 0.30, 0.10
for i, (s, e) in enumerate(keep):
    gap_next = keep[i+1][0] - e if i+1 < len(keep) else TOTAL - e
    gap_prev = s - keep[i-1][1] if i > 0 else s
    add_end   = min(END_BUF,   max(0, gap_next * 0.5))
    add_start = min(START_BUF, max(0, gap_prev * 0.5))
    if gap_next < 0.6: add_end   = max(add_end,   gap_next - 0.05)  # eat tiny gaps
    if gap_prev < 0.6: add_start = max(add_start, gap_prev - 0.05)
    final.append([s - add_start, e + add_end])

The result: cut "find the tickets" extends from src 08:09.26 to 08:09.56 — the trailing 's' survives. Without this, the user will tell you the audio "feels rushed" or "swallows endings".

Rule 5 — Manual duplicate-take removal

Programmatic dedup of identical n-grams is fragile and dangerous (it can drop legitimate repetitions). Instead: read the transcript, find verbatim duplicates yourself, and add explicit drop ranges in source-time:

DUP_RANGES = [
    (78.86, 113.82),   # 01:18.86 - 01:53.82: 2 of 3 takes of "skills as you work"
]
for ds, de in DUP_RANGES:
    # trim each keep that overlaps this range
    ...

Reduces the cut from 35% saved to 32% saved (the duplicates are silenced gaps + speech), but the result reads cleanly.

Output

The script writes:

• cuts.json — keep ranges + cut ranges + summary stats + source-time mapping (used by Phase 8 for overlay anchors).
• edl.txt — human-readable EDL with transcript snippets per keep (eyeball-check this before rendering).

See scripts/build_cuts.py for the full implementation.

---

Phase 5 — Scrub transient beeps

OBS recordings often contain short transient artifacts — boot beeps, key-switch clicks, mic plosives that hit a hot input. These ride along through silence-cut decisions because they're loud (so silence detection ignores them) and they're surrounded by quiet. They sound like a "tick" in the final video at 4:11 or 0:52 or wherever.

Don't try to fix this with a global noise gate or de-noiser. You'll smear consonants and make speech sound underwater. Instead: detect transients and surgically silence each 60–100ms window.

Detection algorithm (the one that doesn't mute speech)

A transient is a "beep" iff:

RMS(window 5ms)  > BEEP_RMS_MIN     (e.g., > 800)
RMS(preceding 150ms) < PRE_QUIET    (e.g., < 200)
RMS(following 200ms) < POST_QUIET   (e.g., < 350)

The third condition is critical. Without it, you'll catch consonant onsets ('p', 't', 'k') of speech, because they look identical in the first 5ms — but speech is followed by a sustained vowel, while a beep returns to silence. Filter on post_avg < 350 and you only ever silence isolated transients.

Also cap the silenced span at 0.30s — a misclassification can never wipe out a word.

Apply the surgical mutes

Build an ffmpeg volume filter expression with the targeted ranges:

volume=enable='between(t,52.110,52.180)+between(t,169.850,169.920)+between(t,251.245,251.315)+between(t,268.780,268.845)':volume=0,adeclick

The trailing adeclick catches anything else click-shaped that the surgical pass missed. Keep both — they're complementary.

See scripts/scrub_beeps.py for the full implementation.

Why per-clip fade-in DOESN'T work

You'd think: "just fade in the start of every kept clip and the boundary clicks disappear." That works for boundary clicks (con​cat artifacts). It does NOT work for transients sitting deeper inside the clip — e.g., 0.7s into the segment, in a silence. We had a beep at cut 4:11.245 that was 0.74s into keep[26] — no fade-in reaches that.

The right fix is the right tool for each problem:

Problem	Fix
Boundary click between two re-encoded clips	Use con​cat filter, not con​cat demuxer (Phase 6)
OBS start tick at t=0	0.4s afade=t=in:st=0:d=0.4
Click-type spikes anywhere	adeclick filter (broad, low cost, no audible side effects)
Transient pip in a silence (the hard one)	Detection + surgical volume=0:enable='between(t,a,b)'

---

Phase 6 — Render the master via con​cat filter

The single most important rendering rule: use the con​cat filter, not the con​cat demuxer.

Why

The con​cat demuxer (-f con​cat -i list.txt) requires every clip to have identical codec params. Even when they do, the demuxer can introduce audible boundary clicks because it stitches packets at container level without aligning audio samples. We saw clicks at every clip join — sounded like the OBS tick we tried to remove, which is what made the user think we hadn't fixed the beep.

The con​cat filter (filter_complex … con​cat=n=N:v=1:a=1) does sample-accurate con​catenation through the audio path. No boundary clicks. Single pass.

The full ffmpeg command

# Build filter_complex string from cuts.json keeps
python3 - <<'PY' > filter_complex.txt
import json
keep = json.load(open('cuts.json'))['keep']
total_out = sum(e-s for s,e in keep)
parts = []
for i,(s,e) in enumerate(keep):
    parts.append(f"[0:v]trim=start={s:.4f}:end={e:.4f},setpts=PTS-STARTPTS[v{i}]")
    parts.append(f"[0:a]atrim=start={s:.4f}:end={e:.4f},asetpts=PTS-STARTPTS[a{i}]")
inter = "".join(f"[v{i}][a{i}]" for i in range(len(keep)))
parts.append(f"{inter}concat=n={len(keep)}:v=1:a=1[vc][ac0]")
parts.append(f"[ac0]adeclick,afade=t=in:st=0:d=0.4,afade=t=out:st={total_out-0.5:.2f}:d=0.5[ac]")
print(";".join(parts))
PY

ffmpeg -hide_banner -y -i "$SRC" \
  -/filter_complex filter_complex.txt \
  -map "[vc]" -map "[ac]" \
  -c:v libx264 -preset medium -crf 18 -pix_fmt yuv420p \
  -c:a aac -b:a 192k -ar 48000 \
  -movflags +faststart \
  master.mp4

Rendering checks

After rendering, verify both streams have matching durations:

ffprobe -v error -show_entries stream=codec_type,duration,nb_frames \
  -show_entries format=duration -of json master.mp4

Video and audio durations should match within 0.05s. If they don't, your filter_complex has a bug — usually because the audio passed through aselect instead of atrim and got time-shifted. Use atrim, not aselect, for cut-list rendering.

After surgical beep mute (Phase 5), the audio has been re-encoded — copy the video stream, replace the audio:

ffmpeg -y -i master.mp4 -c:v copy -af "volume=enable='...':volume=0,adeclick" \
  -c:a aac -b:a 192k -ar 48000 master_clean.mp4

---

Phase 7 — Generate hero imagery

For dramatic full-takeover scenes (the "Apple keynote moment") you want a single hero image — something the audience will remember. Don't use stock photos; generate one with Gemini 3 Pro Image.

Prompt anatomy that works

A reliable Apple-keynote prompt has these 6 elements:

1. Subject + framing — "exploded view of an Apple Watch, components floating in 3D space"
2. Lighting — "soft volumetric studio lighting, specular highlights, rim light"
3. Background — "dark gradient (deep navy to black, subtle teal rim glow)"
4. Style — "photorealistic ray-traced metals and glass, ultra sharp"
5. Layout — "front-3/4 angle, cinematic 16:9 keynote slide composition, generous negative space, no text"
6. Mood — "premium product photography"

The "no text" is critical — image-gen models love adding gibberish text. Always say no.

API call

curl -s -X POST \
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-3-pro-image-preview:generateContent?key=$GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{"parts": [{"text": "<your prompt>"}]}],
    "generationConfig": {
      "responseModalities": ["TEXT", "IMAGE"],
      "imageConfig": {"aspectRatio": "16:9"}
    }
  }' > resp.json

Decode the inline-base64 image:

import json, base64
d = json.load(open('resp.json'))
for p in d['candidates'][0]['content']['parts']:
    if 'inlineData' in p:
        open('hero.png','wb').write(base64.b64decode(p['inlineData']['data']))

When you can't use Gemini

If GEMINI_API_KEY isn't available, fall back to:

• Anton's project: paid-tier EasyChamp Gemini key (per ~/.zshrc) — see reference_gemini_api_keys.md for the user's existing keys
• Free fallback: Pollinations.ai (lower quality, but free)
• Don't ask the user for an API key — check their environment first

Keep the resulting PNG in assets/. Reference it from the composition with <img src="assets/hero.png">.

---

Phase 8 — Author the HyperFrames composition

This is where most visual quality lives. The Layer 2 lessons here are dense — read them all before writing HTML.

8.1 — The iOS-card overlay system (NOT the PiP layout)

The mistake: shrinking the underlying video to a corner PiP and filling the rest with graphics. The lesson: real product videos keep the speaker visible at full screen and float overlays on top like iOS notification cards.

Why "PiP corner" feels wrong

A PiP shrinks the speaker to ~480×270 in a corner. Three problems:

1. The user can no longer see what they're showing on screen at full detail.
2. PiP transitions are jarring — content snaps from full-screen to thumbnail and back.
3. It looks like a Zoom call, not a product video.

The iOS card pattern that works

The card is a frosted-glass surface that drops in on top of the full-screen video. The video keeps playing. The card stays for 4–11 seconds, then dismisses.

.card {
  position: absolute;
  background: linear-gradient(180deg, rgba(28,30,38,0.82) 0%, rgba(18,20,26,0.82) 100%);
  backdrop-filter: blur(28px) saturate(140%);
  -webkit-backdrop-filter: blur(28px) saturate(140%);
  border: 1px solid rgba(255,255,255,0.10);
  border-radius: 28px;
  box-shadow:
    0 1px 0 rgba(255,255,255,0.06) inset,
    0 24px 60px rgba(0,0,0,0.55),
    0 8px 16px rgba(0,0,0,0.35);
  overflow: hidden;
}
.card::before {
  /* Top accent stripe — Skill Studio FEATURES card pattern */
  content: ""; position: absolute; left: 0; right: 0; top: 0;
  height: 4px; background: var(--card-accent, #06b6d4);
  opacity: 0.85;
}

The four ingredients:

1. backdrop-filter: blur(28px) saturate(140%) — the frosted-glass effect that makes the card legible against any underlying video frame. saturate(140%) keeps colors punchy through the blur.
2. Tinted gradient surface — not pure black. rgba(28,30,38,0.82) lets enough of the video bleed through to feel layered, not pasted-on.
3. Inset highlight + drop shadow stack — the 0 1px 0 inset rim catches light, the doubled drop shadow gives depth without looking heavy.
4. Top accent stripe — 4px colored bar tied to the scene's brand color. Single visual anchor.

Card placement — avoid the speaker's face cam

Find where the face cam lives in the source. For most OBS configs, it's bottom-right at ~480×270. Place cards anywhere else:

Card position	When to use
Top, full width	Quote opener, eyebrow + title scenes
Top-left	Big-number stats
Center	Hero brand reveal (one-shot, dramatic)
Bottom-left chip	Small contextual annotations (name corrections)
Bottom, full width	CTA, terminal command

Never place a card in the bottom-right unless the speaker isn't there. Never make a full-screen card cover the screen content unless the moment is specifically designed as a "takeover" beat (e.g., the Apple Watch hero in our project).

8.2 — Real brand icons, not text labels

The project's docs-site/src/components/ui/Icon/brands/ directory ships with 10 brand SVGs:

• claude.svg (#D97757 terra-cotta)
• github.svg (#6e40c9)
• copilot.svg (#6E5494)
• cursor.svg (orange tetrahedron, multi-fill)
• jira.svg (#2684FF)
• azure.svg (#0078D4)
• windsurf.svg, discord.svg, linear.svg, youtube.svg

Copy them into your composition's assets/brands/ and inline them as SVG inside cards, not via <img>. Inlining lets you control the icon's container size and respect the design language. Use the SVGs verbatim — these are the canonical brand marks; do not recolor.

<div class="brand-icon">
  <svg viewBox="0 0 24 24" fill="#D97757" xmlns="http://www.w3.org/2000/svg">
    <path d="…"/>
  </svg>
</div>

.brand-icon { width: 64px; height: 64px; flex: 0 0 64px;
  display: inline-flex; align-items: center; justify-content: center;
  border-radius: 16px;
  background: rgba(255,255,255,0.06);
  border: 1px solid rgba(255,255,255,0.10);
}
.brand-icon svg { width: 38px; height: 38px; }

8.3 — Mirror Skill Studio's tinted-card pattern

Skill Studio (vskill-platform) uses this pattern for feature cards. Our overlays should match:

// Each scene gets a brand color; that color drives the entire card's tint
const SCENE_COLORS = {
  opener:    "#06b6d4",   // cyan
  problem:   "#ef4444",   // red
  solution:  "#22c55e",   // green
  cta:       "#22c55e",
  brand:     "#a855f7",   // purple
  info:      "#3b82f6",   // blue
};

Pill labels follow the exact pattern from vskill-platform/src/app/studio/page.tsx:

.pill {
  display: inline-flex; align-items: center; gap: 8px;
  padding: 6px 14px; border-radius: 999px;
  font-family: 'Geist Mono', 'SF Mono', monospace;
  font-size: 16px; font-weight: 600;
  letter-spacing: 1.5px; text-transform: uppercase;
  /* Per-color: */
  color: var(--c);
  background: rgba(<c-rgb>, 0.10);   /* ${color}10 */
  border: 1px solid rgba(<c-rgb>, 0.40);  /* ${color}40 */
}
.pill .swatch {
  width: 8px; height: 8px; border-radius: 50%;
  background: var(--c);
  box-shadow: 0 0 10px var(--c);
}

This is the exact pattern from the platform — borders at 40% alpha, surfaces at 10% alpha, body text at full color. Mirror it for overlay legibility against any video frame.

8.4 — Sharp animations, not "fade in"

Plain opacity 0 → 1 looks flat. Every card uses a four-property entry:

function cardIn(sel, start) {
  tl.fromTo(sel,
    { opacity: 0, y: 60, scale: 0.96, filter: "blur(8px)" },
    { opacity: 1, y: 0, scale: 1, filter: "blur(0px)",
      duration: 0.7, ease: "power3.out", overwrite: "auto" }, start);
}

The four properties layered:

• Opacity — 0 → 1
• Y offset — 60 → 0 (slide up from below)
• Scale — 0.96 → 1 (subtle pop)
• Blur — 8px → 0 (focus-in)

Plus per-element flourishes:

Element	Animation	Why
Stat numbers	scale 0.5 → 1 with back.out(1.5)	Big-number "Apple keynote pop"
Tool icons	stagger 0.08s with back.out(1.4)	Each one snaps in
Brand wordmark	scale 0.7 → 1 with blur(20px) → 0	Reveal moment
Live dot, terminal cursor	@keyframes blink 1.4s	Lived-in feel
Hero image (Apple Watch)	Continuous gentle y bob 4s sine yoyo	Floats in space

Always use overwrite: "auto" on tweens that share a property — the lint catches it otherwise.

8.5 — Source-time → cut-time mapping (the silent killer)

You wrote overlay anchors against the original source timestamps ("when he says 'SpecWeave' at src 03:28"). The cut master plays at cut timestamps (the same speech moment is at cut 02:19.5).

After every re-cut, the cumulative durations shift slightly. Recompute the mapping and re-resolve every overlay anchor before rendering:

mapping = []
cum = 0.0
for s, e in keep_ranges:
    mapping.append({"src_start": s, "src_end": e, "cut_start": cum})
    cum += (e - s)

def src_to_cut(src_t):
    for m in mapping:
        if m['src_start'] <= src_t <= m['src_end']:
            return m['cut_start'] + (src_t - m['src_start'])
    return None  # source moment was cut

Skip overlays whose source moments got cut — don't try to "place them somewhere reasonable." Drop them.

Save the mapping to cuts.json and have the overlay-author script read it. See scripts/build_cuts.py for the full version.

8.6 — Concise text, big and structured

Cards follow the eyebrow-pill + headline + structured content pattern (Apple keynote):

┌────────────────────────────────────────┐
│ ● PILL LABEL                           │
│                                        │
│ The Big Headline.                      │
│                                        │
│ Tool · Tool · Tool   (or: 3 cards)     │
└────────────────────────────────────────┘

Rules:

• Headline: 56–96px, font-weight: 800, letter-spacing: -1.5px. One sentence. No more.
• Eyebrow pill: 16px mono, all-caps, tracked +1.5px. ≤ 4 words.
• Body: short structured units — bullets (3 max), tool tiles, stat tiles. Never a paragraph.
• Numbers: huge. 140–320px is normal. Numbers are the visual anchor of every stat scene.

Bad: "With more than 600 increments, 500 NPM releases, and over 3000 commits, SpecWeave is battle-tested." Good: ● BATTLE-TESTED / Real numbers. / 3 stat tiles: 600+ / 500+ / 3K+

---

Phase 9 — Lint + verify in Studio

9.1 — HyperFrames lint gotchas

Run npx hyperframes lint after every edit. The five errors/warnings we hit and how to fix:

Lint message	Cause	Fix
overlapping_clips_same_track	Two data-track-index clips on the same track overlap in time	Move one to a different track. Audio especially — give it data-track-index="-1".
overlapping_gsap_tweens	Two GSAP tweens animate the same property on the same selector	Add overwrite: "auto" to all but the first.
gsap_exit_missing_hard_kill	Exit fade ends near a clip start boundary; non-linear seek can land mid-fade	After the exit, add tl.set(sel, { opacity: 0 }, end_time);
composition_file_too_large	Single index.html exceeds ~500 lines	Move scenes into compositions/sceneN.html and mount with data-composition-src. Cosmetic when timing is fluid; do it once timing is locked.
timeline_track_too_dense	More than ~6 elements on one track	Same fix as above — split into sub-comps.

9.2 — Studio gotchas

Symptom	Cause	Fix
New audio plays correctly but waveform display still shows old version	.waveform-cache/ holds stale per-asset JSON	rm -rf .waveform-cache/ then refresh the studio.
Frame-jump number doesn't match expected seconds	Studio internally runs at 30fps regardless of source frame rate	Multiply seconds × 30 to get the studio frame number.
<audio> element conflicts with <video> on same track	Both clips on data-track-index="0" overlap	Audio gets data-track-index="-1" (background, never visible).
Hot-reload doesn't pick up new asset	The file replaced has the same name as before	Refresh the browser tab. Studio polls but is occasionally lazy.
npm run dev fails on second run with port conflict	Previous studio still running on port 3002	lsof -nP -iTCP:3002 then kill the process.

9.3 — Verify scenes by jumping to known cut times

The studio has an "Jump to frame" input. Compute the frame for each scene:

scenes = {
  "quote_opener":      0.5,   # cut 0:00.5
  "boris_correction":  10.4,
  "brownfield":        53.88,
  "apple_watch":       91.4,
  "specweave_reveal":  139.51,
  ...
}
for name, cut_sec in scenes.items():
    print(f"{name}: frame {int(cut_sec * 30)}")

Jump to a frame just past each scene's start (so the entry animation has settled), screenshot, eyeball, iterate.

---

Reference: full pipeline scripts

Three Python scripts handle the heavy lifting. Read them, copy them, adapt them. Don't reinvent.

• scripts/build_cuts.py — Phase 4: smart cut list with word-boundary buffers + dup removal + source→cut mapping.
• scripts/scrub_beeps.py — Phase 5: detect transient beeps (loud + quiet on both sides) + emit ffmpeg volume-gate filter.
• scripts/render_master.py — Phase 6: build the con​cat-filter filter_complex and run ffmpeg single-pass.

Reference: HTML composition starter

• references/composition-starter.html — minimal HyperFrames composition with the iOS-card primitives, brand-icon container, pill labels, and the standard cardIn/cardOut GSAP helpers. Drop in your scene content; ship.

---

Anti-patterns — never do these

1. Don't shrink the video to a PiP corner during overlays. Float overlays on top instead. The user said "those animations must be on top, like phone hover style" — they meant it.
2. Don't fade between cuts with a slow crossfade. Hard cuts (snapped to word edges with buffers) read as professional. Crossfades read as amateur.
3. Don't use auto-editor to do silence cuts. It's choppy, doesn't snap to word boundaries, doesn't handle duplicate takes, doesn't scrub beeps. Use the smart-cut algorithm.
4. Don't use the con​cat demuxer for re-encoded clips. Use the con​cat filter — it sample-aligns audio. Boundary clicks are not OBS ticks; they're con​cat artifacts.
5. Don't fade-in every clip to suppress beeps. Beeps inside silences (deep into a clip) won't be reached. Use surgical detection + targeted mute instead.
6. Don't write paragraphs in cards. Headline + 3 bullets max. If you need more, you need more cards.
7. Don't reinvent brand marks. Inline the official SVGs from docs-site/src/components/ui/Icon/brands/. Don't ask Gemini to "draw a Claude logo".
8. Don't author overlays before the cut master is finalized. Every recut shifts cut times — you'll waste hours re-resolving anchors. Lock the cut, then build overlays.
9. Don't trust Whisper on brand names. "Claude Code at Anthropic" → "Cloud Code at Infropic". Read the transcript and add overlay corrections. Don't re-record.
10. Don't render with large-v3-turbo Whisper on CPU. 15+ minutes for an 8-minute audio. base.en is 30 seconds and equally good for English tech monologue word-boundary work.

---

Verification checklist

Before declaring the video ready:

• Total duration matches the cut summary (within 0.05s).
• Video and audio stream durations are equal in ffprobe output.
• No mid-word clips on the EDL transcript snippets — read each keep's snippet end-to-end.
• Surgical beep ranges all 60–300ms, never longer.
• At every kept-segment boundary, listen for the trailing consonant on the previous segment and the leading consonant on the next.
• Lint reports 0 errors (warnings are fine — they're cosmetic).
• Each overlay scene has been screenshotted in the studio and visually checked for coverage of the speaker face cam.
• Brand icons render correctly (inline SVG, not broken <img>).
• Hero takeover scene (if any) has the underlying video opacity animated to 0 during the takeover, then back to 1 after.
• Final render plays end-to-end without audio glitches; specifically check the cut times where you previously heard beeps.

---

Glossary

• Smart cut — keep ranges built from silence + scene-change correlation, not just silence. Dead air with screen activity is kept; dead air without it is cut.
• Word-boundary buffer — extending a keep range past the word's reported end to preserve trailing consonants (Whisper times vowels, not tails).
• Surgical mute — silencing a sub-100ms window via volume=0:enable='between(t,a,b)', applied only to detected transients in dead air.
• Hero takeover — a full-screen graphic scene that briefly replaces the video entirely (e.g., Apple Watch), then returns. Use sparingly — once or twice per video.
• iOS card — frosted-glass overlay primitive (backdrop-blur + tinted gradient + inset highlight + drop shadow) that floats on top of full-screen video without shrinking it.

---

Provenance

This skill encodes lessons from one specific project: turning Anton Abyzov's 8:47 SpecWeave intro recording (May 2026 OBS capture) into a 6:00 production-ready video. The smart-cut algorithm, beep scrubber, and iOS-card design system were iterated through five render passes (v1 PiP layout → v2 word-boundary buffers → v3 con​cat-filter render → v4 surgical beep mute → v5 iOS-card overlay redesign) before reaching production quality. The Apple-Watch hero image was generated via Gemini 3 Pro Image with the exploded-product-photography prompt template.

Author: Anton Abyzov ([email protected]) Iteration log preserved in ~/Projects/Obsidian/personal-docs/ under 004 Archive/specweave-video-2026-05.