Agent-Skills.md

Agent Skills: Motion Review

Use when judging a rendered film (draft or master) from its frames alone — review pack (ffprobe, stills at 0.0/0.5/1.5 s/mid/last, one still per cut, area-weighted OKLCH planes per still, opacity-blend and drift detection, motion energy, cuts, gesture-free runs and gesture onsets per second, genre + look.md copies), hard-fails checked first (Flat F1-F11 incl. value contrast, display type, centered composition, type scale, mid-tone, full bleed, gesture-free span > 0.75 s, under-animated < 1 onset/s; Motion M1-M3 crossfade/drift/zoom yo-yo; Decorative effect D1-D9 incl. decorative blur, bouncy overshoot, ambient or unproven 3D; Legibility L1-L4; Slop X1-X4; Safety S1-S6), 9-criterion rubric (hook, one idea, motion behaviours, tempo grid, CTA hold, accent, phone legibility, sound-off, look.md consistency) with 1/5/10 anchors, symptom → cause table, and the fixed PASS/REVISE/FAIL/UNCERTAIN verdict.

UncategorizedID: fusengine/agents/motion-review

Repository

fusengine/agents
FusengineLicense: MIT
345

Install this agent skill to your local

pnpm dlx add-skill https://github.com/Fusengine/claude-plugins/tree/HEAD/plugins/motion-expert/skills/motion-review

Skill Files

Browse the full folder contents for motion-review.

Download Skill

Loading file tree…

plugins/motion-expert/skills/motion-review/SKILL.md

Skill Metadata

Name
motion-review
Description
"Use when judging a rendered film (draft or master) from its frames alone — review pack (ffprobe, stills at 0.0/0.5/1.5 s/mid/last, one still per cut, area-weighted OKLCH planes per still, opacity-blend and drift detection, motion energy, cuts, gesture-free runs and gesture onsets per second, genre + look.md copies), hard-fails checked first (Flat F1-F11 incl. value contrast, display type, centered composition, type scale, mid-tone, full bleed, gesture-free span > 0.75 s, under-animated < 1 onset/s; Motion M1-M3 crossfade/drift/zoom yo-yo; Decorative effect D1-D9 incl. decorative blur, bouncy overshoot, ambient or unproven 3D; Legibility L1-L4; Slop X1-X4; Safety S1-S6), 9-criterion rubric (hook, one idea, motion behaviours, tempo grid, CTA hold, accent, phone legibility, sound-off, look.md consistency) with 1/5/10 anchors, symptom → cause table, and the fixed PASS/REVISE/FAIL/UNCERTAIN verdict."
<objective> The judging half of the fuse-motion pipeline. `motion-critic` uses it on every draft (before owner gate G2) and on the master (before the "done" claim). The critic sees what a stranger scrolling a feed would see: frames. It never sees the brief's message or the scene source. The only non-frame inputs are the `genre:` line of the brief and the locked look (`look.md`, `refs.md`, `board.png` from `motion-art-direction`), copied into the pack.

Richness comes from value contrast, type as a large object, composition and rhythm. Flatness is a hard-fail, not a low score. A lone paper plane with small objects, display type under 180 px, a TYPE_SCALE under 4, a centered hero, mid-tone fills, too few full-bleed fields, a span over 0.75 s with no gesture (final CTA still over 1.2 s), fewer than 1 gesture onset per second, a drift and a crossfade each fail the film. So does any decorative effect, with no exception (the bounded swap blur, master motion blur, ≤ 3 % overshoot and purposeful 3D are not effects). Hard-fails come first, then 9 scored criteria. PASS = no hard-fail, every scored criterion ≥ 7 and a mean ≥ 8.0. ffmpeg missing → UNCERTAIN. One pass only. </objective>

Motion Review

Agent Workflow (MANDATORY)

  1. Builder runs bash .motion/review-pack.sh <video> (template below) and hands the critic only: video path, .motion/review/, byte budget, expected pix_fmt.
  2. Critic reads ffprobe.json first; ffprobe/ffmpeg missing → UNCERTAIN, stop. Then genre.txt (missing → assumed ad).
  3. Restate the message in one sentence from the frames — before opening the rubric.
  4. Hard-fails first (hard-fails.md): confirm or reject each flags.txt candidate on frames with the critic-safe checks (review-pack.md). Check every F, M, D, L, X (by genre) and S row.
  5. Read the look (look.md, refs.md, board.png in the pack), then score the 9 criteria (rubric.md); map weak scores via symptom-cause.md.
  6. Emit the block in verdict-format.md — hard-fails listed first. Stop.

Overview

| Step | Input | Output | Reference | |------|-------|--------|-----------| | Pack | video + .motion/brief.md (genre line), look.md, refs/ | facts, stills, beats, planes, mix, drift, energy, cuts, freezes, flags, copies | review-pack.md | | Gate | pack | F / M / D / L / X / S list | hard-fails.md | | Score | pack + look copies | 9 scores 1-10 with frame refs | rubric.md | | Diagnose | scores < 7, hard-fails | symptom → cause → fix rows | symptom-cause.md | | Decide | all of the above | one verdict block | verdict-format.md |

Critical Rules

  1. Intake = the pack — frames, pack text files, genre.txt and the look copies. Never the brief's body, beat map, facts, scene code or builder notes; a leak is discarded and reported as intake_leak: yes.
  2. Message before rubric — no one-sentence restatement → FAIL.
  3. Hard-fails decide first — any F, M, D, L, X (ad) or S row → FAIL whatever the mean.
  4. Flatness and decoration are never scored — F2-F10 and D1-D9 are hard-fails even when look.md asks for them; floors equal the look.md dials, in px at 1080 wide and OKLCH L.
  5. Evidence per score — every score and finding cites a timestamp or pack image.
  6. PASS is strict — no hard-fail, every scored criterion ≥ 7, mean ≥ 8.0; never rounded up.
  7. One pass, no code — name symptom, cause and fix direction; the builder fixes.

Architecture

.motion/
├── brief.md                            ← motion-method (only its genre: line reaches the pack)
├── look.md  refs/{board.png,refs.md}   ← motion-art-direction (locked before scene code)
├── review-pack.sh + review-pack-lib.sh ← review-pack-script.md (builder runs it)
├── out/draft.mp4 | master.mp4
└── review/                             (generated, gitignore)
    ├── ffprobe.json  genre.txt         facts — read first
    ├── still-0.0|0.5|1.5|mid|last.png  hook window, middle, end card
    ├── beat-NN.png                     one still per state between cuts
    ├── contact.png  strip-01..12.png  strip.png  phone-390.png
    ├── planes.txt                      per still: planes, vc, field, mid, cx, bleed (F2 F4 F7 F9 F10)
    ├── mix.txt  drift.txt              opacity blends (M1, F1); sub-threshold motion (M2)
    ├── diff.txt  spikes.txt  spread.txt  energy.txt   motion energy, jumps, luma spread
    ├── cuts.txt  segments.txt          cut timestamps; state durations (tempo, holds, CTA)
    ├── stats.txt  freeze.txt  flags.txt  raw metadata; ≥ 0.75 s gesture-free runs; hard-fail candidates
    ├── gestures.txt                    gesture onsets + rate per second (F11)
    ├── look.md  refs.md  board.png     copies the sandboxed critic may read
    └── frames.txt                      image → frame → timestamp

→ See review-pack-script.md for the complete script

Reference Guide

Concepts

| Topic | Reference | When to Consult | |-------|-----------|-----------------| | Review pack | review-pack.md | Reading the pack; critic sandbox; critic-safe checks (mosaic, halves, psnr, band) | | Hard-fails | hard-fails.md | Every pass, before scoring — the only place IDs are defined | | Worked example | hard-fails-example.md | Calibrating on the rejected flat ad, before and after a cosmetic patch | | Rubric | rubric.md | Scoring — after the restatement; genre rules; audio | | Symptom → cause | symptom-cause.md | Every score < 7 and every hard-fail | | Verdict format | verdict-format.md | Writing the output |

Templates

| Template | When to Use | |----------|-------------| | review-pack-script.md | .motion/review-pack.sh — builder runs it on draft and master |

Best Practices

DO

  • Read ffprobe.json, genre.txt, then flags.txt, before any image
  • Judge the hook on still-0.0, still-0.5, still-1.5; judge each state on its beat-NN
  • Cite the dial (VALUE_CONTRAST, TYPE_SCALE, display 180 px, HOLD) on every F line
  • Confirm F2/F9/F10 on a mosaic, F7 on the halves, F6/M2 with psnr, F3/F8/L2 with band tests
  • Keep the findings table between rounds so a repeated row is visible

DON'T

  • Pass a white-ground film with small centered blocks because its lightest-to-darkest spread is high: F2 weighs areas
  • Credit any gradient background, decorative blur, glow, shadow, grain, glass/chrome, ambient 3D, particle or bouncy overshoot — it is a D fail
  • Fail 3D for being 3D: only decorative/ambient 3D, a banned material/effect, or unproven lit determinism is D7
  • Accept a film that is readable but under-animated — F6 / F11
  • Accept a fade-up, a crossfade or a slow camera push as "motion" — M1 / M2
  • Trust a flags.txt number without the frames, or dismiss a flat film the flags missed
  • Soften a verdict on request or run a second pass on the same file