Malik-rgb43/editing-workflow

edit-motion-graphics

Outcome owner for a finished motion piece (MP4) built as HyperFrames HTML + one paused GSAP timeline. The frame-level PROMPT.md is the product; the code only renders it.

Non-negotiable rules (read first)

1. No code before an approved frame-level PROMPT.md - also in autonomous runs: the agent drafts, a human approves (decision default Q6). Approval is a recorded line APPROVAL: <who> <date> in the file. Change order is always ledger -> PROMPT.md -> code.
2. No spend here. AI-3D, AI stills or any paid generation go through paid-generation-gate; this skill makes no paid call. A catalogue listing is not account access.
3. Review order: Studio preview -> hf_segment ranges -> ONE full render per round of notes -> every-frame QA on the final render. One heavy job at a time (render_lock). Batch the notes; diagnose while the user is still writing, but do not render.
4. Hard traps: no dir="rtl" on the HyperFrames root (RTL on text elements only, lang="he" on <html>); fonts from files (@font-face in hf/fonts/); hyperframes snapshot --describe false, <= 5 timestamps per call; ASCII work root (new_project).
5. Fail closed: a gate that cannot run is blocked (not_run), never pass. A timeout, an empty sample or a missing render never passes. A successful tool call is execution evidence; a viewed render is appearance evidence; a still cannot prove motion.
6. Law vs taste: rules 1-5 are law. Items marked studio style below (no captions by default, banned list, "no pink", Higgsfield grammar, 21st-first) are the owner's taste, taught with their reason; a brief that says otherwise wins on taste, never on gates.
7. Numbers carry their fps (30 unless the brief says otherwise) and measured timings carry the machine (owner's Windows 11 / Ryzen 7 7435HS / RX 7600S / 32 GB). NVIDIA/Apple timings are unmeasured. Projects restricting spend, installs or publication override this skill.

Inputs -> outputs

In: brief/ledger (from video-intake), brand kit, UI screenshots or components, music/VO. Out: hf/PROMPT.md (frame-level), hf/DESIGN.md (one locked palette + type + motion tokens), hf/cues.js (single timing source for picture and sound), hf/SOURCES.md, the composition, mix.wav, the final MP4, QA reports. Handoffs: notes -> revision-round; other ratios -> multi-video-variants (full re-layout, never a crop); delivery gate -> render-qa-deliver; a reference given -> reference-style-transfer first (no reference -> ask which time range is "the style").

Procedure

1. Register + intake. Pick one register (launch | kinetic | logo | explainer | calm) and lock length, ratios, VO, music, deliverable name (references/launch-register-and-pace.md). A 30 s dense launch felt x1.7 too fast; the approved length was 45 s: decide length first.
2. Draft PROMPT.md + DESIGN.md (6 blocks; the camera, seam and event tables; beat cards for any 3D). Run python scripts/motion_spec_check.py hf/PROMPT.md --register <r>; fix; ask for approval.
3. Source beats (references/ui-sources-and-port.md): catalog -> hf-blocks -> shadcn registries (default) -> 21st.dev only with an account -> hand-build; decide 3D per beat (references/three-d-decision.md).
4. Build one index.html + cues.js; ported components re-authored seek-safe. Then hf_preflight --strict, python scripts/seek_safe_scan.py hf/, python scripts/palette_audit.py hf/DESIGN.md hf/ [--png <3D renders>].
5. Approve visually: four stills + a snapshot of every seam + sheet contact sheet; Studio preview; notes batch.
6. Sound (references/sound-and-captions.md): hf_mix --report; brand sound on every brand event; mux and measure the final file.
7. One full render under the lock with an ETA first; frame_qa, motion_qa, hf_deliver; score with agent-content/benchmarks/motion-graphics.rubric.md using a separate critic; present the file, numbered changes, the ledger ticked and the honest gaps.

Register at a glance (details: references/launch-register-and-pace.md)

Register	New screen	Event cadence	Camera	Sound
launch	every 2-6 s, >= 2 depth layers	<= 0.5-1 s; 45-110/min by hand count	never stops (drift 2-4 %/s, push-ins 8-15 %)	music is the SFX; brand sound on every brand event; ring-out
kinetic	every 1-3 s	up to ~200/min (word swaps count)	type is the camera	keyword lands 0..+7 f after it is spoken
logo	one idea, 3-8 s	result visible by 0.4-0.75 s	one move + settle	brand sound on the reveal; 3-step ending
explainer	per idea, 3-6 s	30-110/min	purposeful, to the object explained	VO-led; SFX only on visible change
calm	holds allowed	sparse is fine	slow or locked	one uniform bed + subtle foley; no risers/drops

Per-beat 3D in five lines

Decide per beat, write the reason in PROMPT.md: Blender for real geometry / hero material on screen > 2 s (EEVEE first; measured 1.2-2.25 s/frame at 384-640 px, +1-1.5 min per launch); Three.js for data-driven, instanced, 3D-UI-with-live-text, edit-in-rounds or exact-hex beats, after a 1 s hf_segment AMD gate; AI-3D only behind paid-generation-gate and a licence check; 2.5D (depth_25d, Depth Anything V2 Small, 8.1-9.9 s per 2 s clip) for a still that needs a small move; none for UI/message beats. A 3D object never covers information and reacts to the word.

"Boring / looks AI / static" at a timestamp

Replace the beat concept, do not polish it. Pick from the menu: a voice orb driven by the real VO spectrum; a data-driven build (bars, counters, route globe); the UI rebuilt in code and shot as an object (perspective, depth of field, 150-300 % magnification); a shared-element morph that carries content; a number or word swap in place; a 3D hero landing on the keyword. Never a small corner "feature pill", a flat screenshot with a small push-in, or a dot on black. Add the note to the Banned line and fix it at every occurrence (global rule when it repeats).

Time budget (modelled or measured on the owner's machine; not student measurements)

Target: first presentable draft of a 30-45 s launch in about 3-4 h with 1-2 correction rounds (the first real launch took 10 drafts, 7 critic rounds, 4 owner-note rounds); exactly ONE full render per round. Render reference (HyperFrames 0.8.98, E11, heavy 40 s / 1200 f): 1 worker 132.6 s, 4 workers 92.8 s, 8 workers 88.1 s; browser GPU -12 %; capture is the bottleneck (CPU use only 2-3 of 16 logical cores). Earlier premium renders: 8-14 min typical, 4-21 observed (0.8.79-0.8.93). Always benchmark one unit and post an ETA line before a job over 3 minutes.

Gates

States: pass | fail | blocked | n/a, always with a reason.

Gate	Predicate	Evidence	If false	Owner	Recheck
G0 intake	register, length, ratios, deliverable name, VO/music source, reference answer recorded	hf/BRIEF.md	ask (3-4 questions per round); blocked	video-intake	brief changes
G1 frame-level spec	6 blocks; every beat has frames + px + easing + sound + transition; Banned line; hex palette; APPROVAL: line	motion_spec_check.py exit 0 (2 = approval missing = blocked)	fix spec; stop before code	this skill + human approver	any retime or new beat
G2 clean & smooth	ONE spline per scene, each move >= 1.2 s, reversals >= 1 s and <= 1.3x, text inside frame at every key; seams exit = entry, hero persists, no repeated technique, no fade	tables in PROMPT + checker; motion_qa 0 stutter ranges on the render	re-time / re-path; replace the beat concept rather than polish	this skill	any camera or seam change
G3 density	per register: launch = new screen every 2-6 s, event gaps <= 0.7 s (declared breaths excepted), 45-110 events/min (hand count)	Events table + 3 frame samples per scene	add or replace beats; stretch long moves x1.6 instead of cramming	this skill	after any retime
G4 3D per beat	every 3D beat has a card: route (Blender / Three.js / AI-3D / 2.5D / none), a one-sentence reason, ETA, gate evidence, licence	beat-card table; Three.js rows: a 1 s hf_segment frame that is not blank	re-decide; AI-3D rows need paid-generation-gate	this skill	new beat or asset
G5 palette lock	one palette table; every hex in sources and every 3D render inside it	palette_audit.py exit 0 (+ --png)	recolour / re-render; blocked if no table	this skill	new asset, new render, colour note
G6 UI sourcing	each UI/effect beat has a SOURCES.md row (origin, licence, version, date) and a seek-safe port; no npx shadcn add in a project; no 21st preview/media copied	SOURCES.md + scan	swap source or hand-build with a reason	this skill	new component
G7 seek-safe	hf_preflight --strict 0 errors; seek_safe_scan.py 0 errors; snapshots at non-sequential times agree	reports + snapshot pair	fix; never "fix" by suppressing without a reason	this skill	after every code patch
G8 sound	brand sound on every brand event; SFX under VO within -18..-26 dB and 1-3 f before picture; music rings out; final file -14 LUFS +-0.5, TP <= -1 (house preset v1, Q5)	hf_mix --report, hf_deliver verify on the FINAL file	re-mix and remux (no re-render)	render-qa-deliver measures	any audio or picture retime
G9 captions	no caption elements unless the brief asks	grep of ids/classes + <direction> line	remove / ask in one line	this skill	brief change
G10 render + QA	ONE full render this round; frame_qa 0 flags; rubric average >= 4.0, no dimension < 3, M1/M3/M7 >= 3; every ledger line ticked	QA reports + critic output	back to the failing gate; max one correction round before escalating	render-qa-deliver	each full render

Recording gate results (hf/QA.md, one line per gate)

G1 pass   motion_spec_check exit 0, 24 beats, approval "<who> 2026-10-02"  (re-run after r2 retime: pass)
G5 blocked  3D render coin_r2.png not readable by palette_audit (not_run) - install Pillow or downscale
G7 fail   seek_safe_scan SS02 index.html:211 Math.random -> seeded hash; rescan pending

A gate with no line is not_run. Never write pass from memory of an earlier round: a retime, a new asset or a patch invalidates it (see the Recheck column).

Presenting a draft (render-qa-deliver owns the delivery gate)

Message = the file at once + numbered changes (by the user's own note numbers) + the ledger with every line ticked + the rubric score from the separate critic + the gaps (what was NOT verified: Hebrew proofread, device overlay, loudness on a phone) + the Studio link. Never present with a flagged frame, a failed gate or an unticked ledger line.

References (load when)

• references/seam-camera-event-tables.md - ALWAYS when writing or reviewing PROMPT.md; also agent-content/techniques/frame-spec-prompt.md and clean-smooth-motion.md.
• references/launch-register-and-pace.md - choosing the register, pace complaints, banned list, easing numbers.
• references/three-d-decision.md - any beat with depth or 3D; agent-content/techniques/2-5d.md for 2.5D.
• references/ui-sources-and-port.md - UI cards, effects, screenshot animation; agent-content/techniques/screenshot-rebuild.md.
• references/seek-safe-and-render-traps.md - writing code, patches, render differences.
• references/palette-lock.md, references/typography-hebrew-kinetic.md, references/sound-and-captions.md - DESIGN.md, Hebrew type, the sound plan.
• references/dated-versions.md - pins and licences (dated; expired = unknown).
• Scripts (python <script> --self-check first): motion_spec_check.py, palette_audit.py, seek_safe_scan.py. They check text and patterns, never appearance.

Stop and escalate

Blocked, not guessed: no approval line; no render possible; a gate input missing; a price or licence unknown; a Three.js beat whose 1 s segment is blank on AMD (fall back to a Blender sprite). Ask before any spend, install, publication or message to a third party. Never message other agent sessions. A user note "boring / looks AI / static" at a timestamp means replace the beat concept from the register's menu, in one line of explanation.

Maintenance

Volatile facts live only in references/dated-versions.md and the global dated modules (hyperframes-traps.md, three-d-routes.md, audio-mix.md). Add every new owner note to the Banned line, not to this file. Specified, deterministic checks only; model eval not run (decision default Q4).