Video Studio (read first; overrides the rest of this file)

This skill is installed in Video Studio (Edits/.claude/skills/), from the NextPhase / creator packs Bryan downloaded on 2026-10-31. Video Studio's CLAUDE.md wins wherever the two disagree.

• Start from the studio job, never an intake. Jobs come from the studio page (jobs/<id>/job.json). This skill runs when the job's style is film-grade. Never ask "short or long?" or "which style?": the page already asked. (nextphase-studio's intake is not used here.)
• One transcript: projects/<id>/transcript/words.json (WhisperX). Don't run this skill's own transcriber.
• Brand: styles/brand.json (purple #8B5CF6 / #C4B5FD / #4C1D95, Inter + Playfair Display) replaces this skill's default colours. Generate its theme/tokens from it.
• Show the motion graphics before rendering (Bryan's rule): design first, show stills of each graphic over the real frame in batches (the design → graphics stages, graphics.json), wait for approval, then render.
• Sounds and music: the studio's library (styles/sfx.json, node studio/sfx.mjs place) on the locked picture; music only from tracks Bryan owns or licensed.
• Windows Git Bash, not macOS/WSL: python not python3; a venv's Python is Scripts/python.exe; long paths fail, so map a short drive (cmd //c "subst R: <winpath>"). Run this skill's setup/doctor script only the first time the style is used, and ask before any download over ~500 MB.
• Where work goes: editing work in projects/<id>/, deliverables in jobs/<id>/.

PERSA — editing style

A talking head plus screen recordings goes in; an edit comes out where the footage itself is the design. There is no dark grid and no lit stage. There is a graded plate, a cut-out person, and type that lives in the room with them.

⚠ Before editing ANY video, ask the user which style: persa (this skill) · nextphase (nextphase-long-editing) · motiongraphics. Never assume, never infer it from the footage. Record the answer at the top of PLAN.md.

Where this look comes from

Reverse-engineered frame-by-frame from three reels Jade supplied. Every colour, size and duration in assets/tokens.css is measured, not invented:

• REF-A — instagram.com/p/DcgMgW6yahM/ — 38.7s, the primary reference.
• REF-B — instagram.com/p/DcOJfx0SXyN/ — 67.6s, same creator, cleaner density.
• REF-C — instagram.com/p/DcBwAjvoEqI/ — a different creator. Used for its font recommendations only (reference/typography.md §4). Its own look is NOT persa; do not mix it in.

reference/reference-log.md has the shot-by-shot breakdown, the measurement method, and a table mapping every finding to where it is written up.

---

⛔ Six hard rules

1. Never hardcode a colour, font or duration. Import assets/tokens.css; use its variables only. Never mix persa tokens with nextphase or motiongraphics tokens in one composition.
2. The word caption HARD-CUTS. Zero frames of transition. No fade, no scale, no pop, ever. This is the single most-broken rule in the style — see reference/captions.md §1.
3. The emphasis word lives BEHIND the subject matte. The head must occlude it. In front, it is a lower-third graphic; behind, it is in the room. That difference is the style.
4. Animate transforms only — opacity, x, y, scale, scaleX/Y, rotation plus filter: blur(). Never top/left/width/height.
5. Every word on screen is derived from THIS video. Kit copy in components/ is placeholder, and some reference copy belongs to other creators. Run bash scripts/check-persa.sh videos/<name> before every render; it exits non-zero on kit copy, borrowed names, and persa-specific style violations.
6. Nothing publishes without explicit approval. Deliver rendered + scored, open it for review, stop there.

Read reference/ before building — grade.md (the measured colour science) · captions.md (the core of the style — read this first) · layouts.md (the five scenes and the matte) · typography.md · motion-grammar.md · sound-design.md · workflow.md · reference-log.md.

---

1 · The look in one paragraph

Blue is pulled down hard everywhere; green is lifted in the shadows and highlights; everything is desaturated, bloomed, vignetted and grained; nothing reaches pure black or pure white. The footage is warm, hazy and slightly old. Over it: one white word at chest height that replaces itself on every syllable, and occasionally one huge outlined word behind the speaker's head. Media arrives as rounded panels the person's shoulders break through. Three accent colours, one per frame. No transitions except a panel fading in place.

Full numbers: reference/grade.md.

---

2 · The frame

1920×1080 @30fps is the default for this studio. The style was measured at 1080×1920, so that is what :root in tokens.css carries; .pa-16x9 overrides the six tokens that do not transfer.

<div class="pa-frame pa-16x9"> … </div>     <!-- long-form -->
<div class="pa-frame">          … </div>     <!-- vertical  -->

Read reference/layouts.md §16:9 before building wide. Three scene rules change, they were derived on real 16:9 footage, and the first one fails silently:

1. A short emphasis word centred behind a centred head disappears. Long words run frame-wide and the head cuts their middle; short words get --pa-emph-x so the head cuts one edge instead.
2. The caption centres on the SUBJECT, not the frame — set --pa-cap-x whenever the speaker sits in a corner card.
3. Media becomes a row of three panels, not one panel stacked above the head.

Zone	Value	Rule
Safe	inset --pa-safe-x 6.5% × --pa-safe-y 4.5%	Nothing outside. Suspended for the subject only.
Caption	centred, 73% height	Fixed. Never moves. Nothing else lives there.
Emphasis	centred, 25% height	Behind the matte. Top two-thirds must stay readable.
Panel	9:16 86% wide, top 8% · 16:9 a row of 3	Scene 2.
Head card	9:16 80% centred · 16:9 34% corner card	Scene 3/4. Bleeds off the bottom.
9:16 bottom	lowest 12%	Keep readable content out — platform UI.

⭐ The z-order — get this wrong and it is not persa

50  .pa-caption   one word, in front of everything
40  .pa-subject   the matted talking head
20  .pa-emph      the huge word the head cuts into
10  .pa-panel     real screen recordings
 0  the graded plate

⭐ .pa-frame on the composition root

Type is sized in cqh so one scale serves both aspects. That resolves against a size container — the root must carry .pa-frame. Without it every string renders at a fraction of its size and the layout still "looks fine", which is exactly why it is worth checking first when a scene looks wrong.

---

3 · The five scenes and the rotation

Full spec in reference/layouts.md.

#	Scene	For
1	Clean Face	the rest beat — and it is ~40% of the runtime
2	Panel Over Face ⭐	the signature: media panel, head breaks its bottom edge
3	Media Full / Head Card	the screen content IS the point; punched in, head card bleeds off the bottom
4	Split / Break the Box	media above, head card below, head crosses the seam
5	Hand-Note Beat	a lettered line, often curved to follow a real edge in the room

Never the same scene twice in a row. Never hold one past ~20s. Every scene change carries a sound. But note what the reference does: it alternates clean face and media beat, and it returns to a clean face constantly, holding 6–10s at a stretch. Wall-to-wall media is the amateur read.

⭐ REAL MEDIA, NOT TYPESET TEXT

Every beat that supports a point shows a real resource — a screen recording, a real product UI, a docs page, a repo card, a real thread. ≥1 per 20s, minimum 8.

• Check every screenshot actually rendered content. Login-walled and JS-heavy pages come back blank and ship as empty panels.
• No random humans as b-roll. Cut to a person only when that person is the subject of the line.
• Never use emojis. Build designed marks — a CSS glyph, a small SVG, a monogram.
• Re-encode every sourced clip to 30fps CFR and verify with ffprobe.

⭐ COPY IS DERIVED, NEVER INHERITED

Every string comes from the speaker's own words at their onset in transcript-tight.json, the real thing named in that beat, or a verified figure. Write the actual copy into PLAN.md, per beat, before authoring — that is where a factual error gets caught cheaply.

---

4 · The two things that make it persa

The grade

python scripts/persa-balance.py assets/recording.mp4 --out assets/bal.mp4  # neutralise first
bash scripts/persa-grade.sh assets/bal.mp4 assets/plate.mp4 plate          # then grade
bash scripts/persa-grade.sh out/visual.mp4 out/visual-look.mp4 soft       # once, at the end
python  scripts/grade-check.py out/visual-look.mp4                        # verify

Grade the plate, composite, then bloom/vignette/grain the finished comp once. Grading footage at full and laying ungraded graphics on top is the most common way this look fails — the graphics sit on the image instead of in it.

Balance before you grade. The persa curve is calibrated against a neutral source; a real plate's own cast rides straight through it, so grading a cool-lit recording directly leaves it failing grade-check.py no matter how warm it looks. persa-balance.py measures the source's cast per luma band and writes the cancelling curve — one command, no hand-tuning. Verified on real footage: four readings out of tolerance before, a clean pass after.

grade-check.py gates on the channel relationships (B-R in the mids is the number that matters — −18 to −28) and passes on both reference reels while failing a third reel by a different creator.

The matte

source ~/.cache/nextphase/rvm_env/bin/activate
python ../nextphase-long-editing/scripts/rvm-matte.py --src assets/plate.mp4 --out assets/matte --frames <n>

Matte the graded plate. Keep the rim — the light fringe on the matte edge is visible in the reference at 3× and is what separates the subject from a busy recording behind them; --pa-rim makes it deliberate instead of leaving it to luck. What you still fix is flicker. Detail: ../nextphase-long-editing/reference/cutout-compositing.md.

---

5 · Motion

reference/motion-grammar.md. The whole cadence, measured at 30fps:

Move	Frames
Word caption	0 — it hard-cuts
Emphasis in	5
Slam	4
Panel in	9
Panel out	8, fading in place — it never slides
Plate drift	continuous, scale 1.00→1.03 over 6s, linear

Persa moves less than either sibling style. There is no whip pan, no zoom blur, no light leak, no glitch. If a cut needs help, the help is a sound.

---

6 · Sound

reference/sound-design.md. Because persa has almost no visual transitions, sound carries more here, not less — it is what makes a hard cut read as intentional. ~10–16 cues/min. Every layout change gets one. Reuse the sibling palette and call the mixer by its path inside that skill folder — never copy it into a video folder:

bash ../nextphase-long-editing/scripts/mix-sfx.sh out/visual-look.mp4 out/sfx.mp4 sfx-cues.txt
bash ../nextphase-long-editing/scripts/mix-bgm.sh out/sfx.mp4 out/final.mp4 assets/music/bed.mp3

Match a click to every real cursor press in a screen recording — that is what makes the panels feel live.

---

7 · The workflow

Full detail in reference/workflow.md. Prefix every bash step with:

export PATH="$HOME/.nvm/versions/node/v24.18.0/bin:$HOME/.local/bin:$PATH"

0. Ask which style. Locate videos/<name>/, VIDEO.md, assets/recording.mp4.
1. Transcribe → chunk anything over ~90s and gap-check minute by minute.
2. Tighten → transcript-tight.json is now the only timing source.
3. Balance (persa-balance.py) → grade the plate (plate strength) → grade-check.py.
4. Matte the graded plate with RVM.
5. PLAN.md — scenes per beat, the actual on-screen copy, the media list, the re-hooks. Get a "go" before authoring.
6. Source real media, verify facts, normalise to 30fps CFR.
7. Author per-chapter compositions. Symlink assets into each seg-NN/; HyperFrames rejects ../assets/. Build captions with scripts/build-captions.py.
8. Per segment: lint (0 errors) → check-persa.sh (must exit 0) → snapshot --at <times> → read the contact sheet and fix → render -q high -f 30.
9. Concat via ../nextphase-long-editing/scripts/concat-segments.sh — chapter timestamps come from real rendered durations, never from PLAN.md. Relay the source audio after the concat (AAC priming drifts ~38ms per join).
10. Finish the look (soft strength over the concat) → grade-check.py.
11. Sound: SFX pass, then the ducked bed.
12. Deliver: out/final.mp4 · out/chapters.txt · out/<name>.srt. Report rendered + scored, open for review. Publish nothing without approval.

---

Files in this skill

• assets/tokens.css — the measured design tokens and every persa class. Import into every composition. (CSS comments do not nest — never write a closing marker inside one of its header blocks, or the whole sheet silently stops applying.)
• components/persa-system.html — the whole style on one page: all five scenes, both caption tiers, the z-order table, the palette and the cadence, live in a browser. Open this first; it is the fastest way to see what the style is. Open it beside reference/stills/.
• reference/ — grade.md · captions.md · layouts.md · typography.md · motion-grammar.md · sound-design.md · workflow.md · reference-log.md.
• reference/stills/ — the source evidence: 14 frames including two 30fps transition strips (the caption hard-cut and the panel exit), 2–4× type crops, a 3× crop of the matte edge, and 1fps contact sheets of all three reels.
• scripts/persa-balance.py — measure a source's cast and cancel it. Run this before grading real footage, or the plate will not land in the colour space.
• scripts/persa-grade.sh — apply the look (plate | soft | full).
• scripts/grade-check.py — verify a clip is in the persa colour space.
• scripts/build-captions.py — transcript → hard-cut word captions.
• scripts/check-persa.sh — the render gate. Run before every render.
• Everything else is shared with nextphase-long-editing — tighten, rvm-matte, mix-sfx, mix-bgm, concat-segments. Call them by their path in that folder; never copy them into a video folder.