[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M72CEE4X4SJWQ4N8RDVCF5V9":3,"skill-same:01M72CEE4X4SJWQ4N8RDVCF5V9":47},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":29,"ownerHtmlUrl":30,"ownerBlogUrl":31,"ownerDescription":6,"updatedAt":32,"install":33,"resources":46},"01M72CEE4X4SJWQ4N8RDVCF5V9","footage-keyframes",null,"Extract scene-aware video keyframes for footage analysis, including cuts, action and edit decisions.","1aGh\u002Fmaude","1aGh","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F17166964?v=4","01M72CAQDBB0YG724KFT7WEYRF","Michal Dovrtel","01M72CAQDXC32AWQRZYNJXXDC4",3,"MIT","https:\u002F\u002Fgithub.com\u002F1aGh\u002Fmaude\u002Fblob\u002Fmain\u002Fplugins\u002Fdesign-video\u002Fskills\u002Ffootage-keyframes\u002FSKILL.md",[],[],"2026-11-10 18:39:20",false,"---\nname: footage-keyframes\ndescription: \"Extract scene-aware video keyframes for footage analysis, including cuts, action and edit decisions.\"\n---\n\n# footage-keyframes — understand the video's dynamics, extract only the frames that matter\n\nFollow [host conventions](..\u002F..\u002FHARNESS.md) for Claude Code or Codex.\n\nBlind frame-rate sampling is the problem this skill exists to fix. `probe-footage`\npulls N evenly-spaced frames; a sub-interval shot (a \u003C0.25 s action flash before an\ninterview, a quick cut) falls **between** samples and is lost — the analyzer then\ncan't see that beat and either misses or hallucinates it. Raising the frame rate\nblindly just inflates cost without guaranteeing the *right* frames land.\n\n`maude design smart-frames` selects frames by the video's **real dynamics**:\nscene cuts + semantic action beats + the true first & last frame. Fewer, meaningful\nframes → sharper analysis at lower cost.\n\n## The three tiers (auto-detected, each a graceful fallback)\n\n| Tier | How it picks frames | Needs | When |\n| ---- | ------------------- | ----- | ---- |\n| **`gemma`** | Gemma **scout** watches the video and flags action beats that are *not* hard cuts (a snap, a run, a reveal inside one continuous shot), merged with ffmpeg scene cuts. Two interchangeable runtimes: **mlx-vlm** (preferred — native video input, Apple Silicon, the benchmarked path) or **Ollama** (simplest install — a gemma3 vision model; frames are sampled by ffmpeg and sent as timestamped images). | ffmpeg **+** (mlx-vlm + a Gemma-4 MLX model, *or* Ollama + `gemma3:4b`+) | opt-in; richest — catches beats inside continuous shots |\n| **`ffmpeg`** | ffmpeg content scene-detection + endpoints + long-shot midpoints. | ffmpeg | **the default** when ffmpeg is present; cross-platform, no model download |\n| **`blind`** | delegates to `probe-footage` (headless Chromium, even-spaced). | Chromium (already shipped) | the floor — so extraction always works even with neither ffmpeg nor Gemma |\n\n`--engine auto` (default) probes availability and degrades **gemma → ffmpeg → blind**.\nAn explicit `--engine X` forces a tier and **errors if its deps are missing** — Maude\nnever silently downgrades an engine the user asked for by name (mirrors the\ntranscription-engine posture).\n\n**No model, no problem.** The whole point of the tiering: a user who doesn't want to\ndownload a multi-GB Gemma model (or isn't on Apple Silicon) gets the `ffmpeg` tier —\nstill scene-aware, just without the semantic-beat layer. A user with nothing gets the\n`blind` floor. The Gemma scout is a bonus, never a requirement.\n\n## Contract — `maude design smart-frames`\n\n```sh\nmaude design smart-frames \u003Casset> [--root \u003Crepo>] [--out-dir DIR] [--frames N]\n                          [--engine auto|gemma|ffmpeg|blind]\n                          [--scene-thresh 0.3] [--scout-fps 4]\n```\n\n- `\u003Casset>` — a content-addressed `assets\u002F\u003Csha8>.\u003Cext>` (production) **or** any readable\n  clip path (terminal power-user analysis of an arbitrary file). The `ffmpeg`\u002F`gemma`\n  tiers accept both; the `blind` tier delegates to `probe-footage`, which requires the\n  content-addressed form.\n- stdout on success = a JSON **manifest** (a strict SUPERSET of `probe-footage`'s, so\n  it's a drop-in for the footage-analyst):\n\n```jsonc\n{ \"asset\": \"assets\u002F\u003Csha8>.mp4\", \"durationSec\": 8.0, \"width\": 1280, \"height\": 720,\n  \"method\": \"ffmpeg\",                       \u002F\u002F which tier actually ran\n  \"scout\": null,                            \u002F\u002F \"mlx\" | \"ollama\" when the gemma scout contributed beats\n  \"sceneCuts\": [0.233, 6.533],              \u002F\u002F ffmpeg scene-detect (empty in blind)\n  \"scoutBeats\": [{ \"t\": 6.9, \"what\": \"run play\" }],  \u002F\u002F gemma only (empty otherwise)\n  \"outDir\": \"$TMPDIR\u002Fsmart-frames-…\",\n  \"frames\": [ { \"index\": 1, \"t\": 0.0, \"png\": \"…\u002Ff_01.png\" }, … ] }\n```\n\nConsumers read `frames[]` (index + source-time `t` + PNG path) exactly as they read\n`probe-footage`'s; `method` \u002F `sceneCuts` \u002F `scoutBeats` are additive hints.\n\n- Exit: `0` ok · `2` usage \u002F asset-not-found · `3` a forced engine's deps are missing\n  · `4` decode\u002Fextract error · `1` other.\n\n## Env knobs\n\n| Env | Default | Effect |\n| --- | ------- | ------ |\n| `MAUDE_SMARTFRAMES_ENGINE` | `auto` | tier override (same as `--engine`) |\n| `SCENE_THRESH` (`--scene-thresh`) | `0.3` | ffmpeg scene-cut sensitivity (lower = more cuts) |\n| `SCOUT_FPS` (`--scout-fps`) | `4` | Gemma scout native sampling density |\n| `--frames` \u002F `--max-frames` | `12` | hard cap on extracted frames |\n| `MAUDE_MLX_PYTHON` | auto (Maude venv → PATH) | a Python that can `import mlx_vlm` (gemma tier, mlx runtime) |\n| `MAUDE_GEMMA_MODEL` | `mlx-community\u002Fgemma-4-e4b-it-4bit` | the scout model (mlx runtime) |\n| `OLLAMA_HOST` | `http:\u002F\u002F127.0.0.1:11434` | the Ollama server (gemma tier, ollama runtime). **Loopback-only** — a non-loopback value is refused, never uploaded to (the footage pipeline is egress-free by design, DDR-183) |\n| `MAUDE_OLLAMA_MODEL` | auto (first vision-capable `gemma3:*` tag) | the scout model (ollama runtime; `gemma3:1b` and `gemma3n` are text-only and skipped) |\n\n## Two front doors, one engine\n\n- **Settings** → the \"Scene-aware keyframes\" section persists the engine\n  choice (`keyframeEngine`) and offers a one-click Gemma-model download (like the\n  Subtitles \u002F whisper section). App-served runs honor that pref.\n- **Terminal \u002F power users** → `--engine auto` **self-detects deps** with no app or\n  config, so `maude design smart-frames \u003Cclip>` is a standalone terminal video\n  analyzer. Neither path requires the other; an explicit `--engine` always wins.\n\n## Consumers (who references this skill)\n\n- **`\u002Fdesign-video:video-analyze`** — the canonical one-shot analysis: the orchestrator runs\n  `smart-frames` + `maude design transcribe-local`, the Read-only `footage-analyst` watches\n  the smart frames (with the transcript folded in) and returns JSON, the orchestrator\n  persists it → a combined **visual + audio** metadata report.\n- **`\u002Fdesign-video:reel`** — runs the **same** shared analysis step (no duplicate workflow)\n  before the director\u002Fcodegen phase.\n- **`footage-analyst`** agent — consumes the `frames[]` this skill produces; it's\n  Read-only\u002Fegress-free (DDR-183 F2), so the orchestrator runs `smart-frames`, not the\n  agent.\n- **`scripts\u002Fvideo-benchmark\u002F`** — the harness this pipeline was proven in; this skill\n  is its productionized form.\n\n## Related\n\n- `maude design transcribe-local` (verb) — the audio half: the words of a clip on this Mac\n  (whisper.cpp), written to `assets\u002F\u003Csha8>.transcript.json`. A cloud engine is `maude design\n  transcribe --provider`, which spends the person's account. Audio quality is\n  engine-independent, so transcription is a separate step from the frames.\n- Implementation: `apps\u002Fstudio\u002Fbin\u002Fsmart-frames.sh` → `_smart-frames.mjs` (pure JS;\n  runs under node or bun). Manifest superset of `apps\u002Fstudio\u002Fbin\u002F_probe-footage-playwright.mjs`.\n","# footage-keyframes — understand the video's dynamics, extract only the frames that matter\n\nFollow [host conventions](..\u002F..\u002FHARNESS.md) for Claude Code or Codex.\n\nBlind frame-rate sampling is the problem this skill exists to fix. `probe-footage`\npulls N evenly-spaced frames; a sub-interval shot (a \u003C0.25 s action flash before an\ninterview, a quick cut) falls **between** samples and is lost — the analyzer then\ncan't see that beat and either misses or hallucinates it. Raising the frame rate\nblindly just inflates cost without guaranteeing the *right* frames land.\n\n`maude design smart-frames` selects frames by the video's **real dynamics**:\nscene cuts + semantic action beats + the true first & last frame. Fewer, meaningful\nframes → sharper analysis at lower cost.\n\n## The three tiers (auto-detected, each a graceful fallback)\n\n| Tier | How it picks frames | Needs | When |\n| ---- | ------------------- | ----- | ---- |\n| **`gemma`** | Gemma **scout** watches the video and flags action beats that are *not* hard cuts (a snap, a run, a reveal inside one continuous shot), merged with ffmpeg scene cuts. Two interchangeable runtimes: **mlx-vlm** (preferred — native video input, Apple Silicon, the benchmarked path) or **Ollama** (simplest install — a gemma3 vision model; frames are sampled by ffmpeg and sent as timestamped images). | ffmpeg **+** (mlx-vlm + a Gemma-4 MLX model, *or* Ollama + `gemma3:4b`+) | opt-in; richest — catches beats inside continuous shots |\n| **`ffmpeg`** | ffmpeg content scene-detection + endpoints + long-shot midpoints. | ffmpeg | **the default** when ffmpeg is present; cross-platform, no model download |\n| **`blind`** | delegates to `probe-footage` (headless Chromium, even-spaced). | Chromium (already shipped) | the floor — so extraction always works even with neither ffmpeg nor Gemma |\n\n`--engine auto` (default) probes availability and degrades **gemma → ffmpeg → blind**.\nAn explicit `--engine X` forces a tier and **errors if its deps are missing** — Maude\nnever silently downgrades an engine the user asked for by name (mirrors the\ntranscription-engine posture).\n\n**No model, no problem.** The whole point of the tiering: a user who doesn't want to\ndownload a multi-GB Gemma model (or isn't on Apple Silicon) gets the `ffmpeg` tier —\nstill scene-aware, just without the semantic-beat layer. A user with nothing gets the\n`blind` floor. The Gemma scout is a bonus, never a requirement.\n\n## Contract — `maude design smart-frames`\n\n```sh\nmaude design smart-frames \u003Casset> [--root \u003Crepo>] [--out-dir DIR] [--frames N]\n                          [--engine auto|gemma|ffmpeg|blind]\n                          [--scene-thresh 0.3] [--scout-fps 4]\n```\n\n- `\u003Casset>` — a content-addressed `assets\u002F\u003Csha8>.\u003Cext>` (production) **or** any readable\n  clip path (terminal power-user analysis of an arbitrary file). The `ffmpeg`\u002F`gemma`\n  tiers accept both; the `blind` tier delegates to `probe-footage`, which requires the\n  content-addressed form.\n- stdout on success = a JSON **manifest** (a strict SUPERSET of `probe-footage`'s, so\n  it's a drop-in for the footage-analyst):\n\n```jsonc\n{ \"asset\": \"assets\u002F\u003Csha8>.mp4\", \"durationSec\": 8.0, \"width\": 1280, \"height\": 720,\n  \"method\": \"ffmpeg\",                       \u002F\u002F which tier actually ran\n  \"scout\": null,                            \u002F\u002F \"mlx\" | \"ollama\" when the gemma scout contributed beats\n  \"sceneCuts\": [0.233, 6.533],              \u002F\u002F ffmpeg scene-detect (empty in blind)\n  \"scoutBeats\": [{ \"t\": 6.9, \"what\": \"run play\" }],  \u002F\u002F gemma only (empty otherwise)\n  \"outDir\": \"$TMPDIR\u002Fsmart-frames-…\",\n  \"frames\": [ { \"index\": 1, \"t\": 0.0, \"png\": \"…\u002Ff_01.png\" }, … ] }\n```\n\nConsumers read `frames[]` (index + source-time `t` + PNG path) exactly as they read\n`probe-footage`'s; `method` \u002F `sceneCuts` \u002F `scoutBeats` are additive hints.\n\n- Exit: `0` ok · `2` usage \u002F asset-not-found · `3` a forced engine's deps are missing\n  · `4` decode\u002Fextract error · `1` other.\n\n## Env knobs\n\n| Env | Default | Effect |\n| --- | ------- | ------ |\n| `MAUDE_SMARTFRAMES_ENGINE` | `auto` | tier override (same as `--engine`) |\n| `SCENE_THRESH` (`--scene-thresh`) | `0.3` | ffmpeg scene-cut sensitivity (lower = more cuts) |\n| `SCOUT_FPS` (`--scout-fps`) | `4` | Gemma scout native sampling density |\n| `--frames` \u002F `--max-frames` | `12` | hard cap on extracted frames |\n| `MAUDE_MLX_PYTHON` | auto (Maude venv → PATH) | a Python that can `import mlx_vlm` (gemma tier, mlx runtime) |\n| `MAUDE_GEMMA_MODEL` | `mlx-community\u002Fgemma-4-e4b-it-4bit` | the scout model (mlx runtime) |\n| `OLLAMA_HOST` | `http:\u002F\u002F127.0.0.1:11434` | the Ollama server (gemma tier, ollama runtime). **Loopback-only** — a non-loopback value is refused, never uploaded to (the footage pipeline is egress-free by design, DDR-183) |\n| `MAUDE_OLLAMA_MODEL` | auto (first vision-capable `gemma3:*` tag) | the scout model (ollama runtime; `gemma3:1b` and `gemma3n` are text-only and skipped) |\n\n## Two front doors, one engine\n\n- **Settings** → the \"Scene-aware keyframes\" section persists the engine\n  choice (`keyframeEngine`) and offers a one-click Gemma-model download (like the\n  Subtitles \u002F whisper section). App-served runs honor that pref.\n- **Terminal \u002F power users** → `--engine auto` **self-detects deps** with no app or\n  config, so `maude design smart-frames \u003Cclip>` is a standalone terminal video\n  analyzer. Neither path requires the other; an explicit `--engine` always wins.\n\n## Consumers (who references this skill)\n\n- **`\u002Fdesign-video:video-analyze`** — the canonical one-shot analysis: the orchestrator runs\n  `smart-frames` + `maude design transcribe-local`, the Read-only `footage-analyst` watches\n  the smart frames (with the transcript folded in) and returns JSON, the orchestrator\n  persists it → a combined **visual + audio** metadata report.\n- **`\u002Fdesign-video:reel`** — runs the **same** shared analysis step (no duplicate workflow)\n  before the director\u002Fcodegen phase.\n- **`footage-analyst`** agent — consumes the `frames[]` this skill produces; it's\n  Read-only\u002Fegress-free (DDR-183 F2), so the orchestrator runs `smart-frames`, not the\n  agent.\n- **`scripts\u002Fvideo-benchmark\u002F`** — the harness this pipeline was proven in; this skill\n  is its productionized form.\n\n## Related\n\n- `maude design transcribe-local` (verb) — the audio half: the words of a clip on this Mac\n  (whisper.cpp), written to `assets\u002F\u003Csha8>.transcript.json`. A cloud engine is `maude design\n  transcribe --provider`, which spends the person's account. Audio quality is\n  engine-independent, so transcription is a separate step from the frames.\n- Implementation: `apps\u002Fstudio\u002Fbin\u002Fsmart-frames.sh` → `_smart-frames.mjs` (pure JS;\n  runs under node or bun). Manifest superset of `apps\u002Fstudio\u002Fbin\u002F_probe-footage-playwright.mjs`.\n",{"name":5,"description":7},"main","cfde9d2b169e4b418a5c5efb35297cc575dc764c","plugins\u002Fdesign-video\u002Fskills\u002Ffootage-keyframes",0,"https:\u002F\u002Fgithub.com\u002F1aGh\u002Fmaude","https:\u002F\u002Fmaude.sh","https:\u002F\u002Fgithub.com\u002F1aGh","www.iagh.cz","2026-11-13 20:38:15",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":45},[35,38,41],{"tool":36,"command":37},"npx","npx skills add 1aGh\u002Fmaude --skill footage-keyframes",{"tool":39,"command":40},"bunx","bunx skills add 1aGh\u002Fmaude --skill footage-keyframes",{"tool":42,"command":43},"pnpm","pnpm dlx skills add 1aGh\u002Fmaude --skill footage-keyframes","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002F1aGh\u002Fmaude\u002Fblob\u002Fmain\u002Fplugins\u002Fdesign-video\u002Fskills\u002Ffootage-keyframes\u002FSKILL.md",true,[],[48,55,62,69,76,83],{"skillNo":49,"name":50,"nameZh":6,"description":51,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":52,"categories":53,"keywords":54,"publishedAt":19},"01M72CJCVNC5JB3M36ES649CTY","workflow-state","Read and update workflow state when planning, executing, completing, pausing or resuming project work.","https:\u002F\u002Fgithub.com\u002F1aGh\u002Fmaude\u002Fblob\u002Fmain\u002Fplugins\u002Fflow\u002Fskills\u002Fworkflow-state\u002FSKILL.md",[],[],{"skillNo":56,"name":57,"nameZh":6,"description":58,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":59,"categories":60,"keywords":61,"publishedAt":19},"01M72CJAJBZT2RHE97ND8EFTA3","ui-rules","UI hard-stop rules and audit method: WCAG 2.1 AA accessibility, responsive layout and motion.","https:\u002F\u002Fgithub.com\u002F1aGh\u002Fmaude\u002Fblob\u002Fmain\u002Fplugins\u002Fflow\u002Fskills\u002Fui-rules\u002FSKILL.md",[],[],{"skillNo":63,"name":64,"nameZh":6,"description":65,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":66,"categories":67,"keywords":68,"publishedAt":19},"01M72CJ857314HCC8A99T4K0D2","testing-rules","Apply test quality rules during implementation and review: meaningful assertions, isolation and coverage.","https:\u002F\u002Fgithub.com\u002F1aGh\u002Fmaude\u002Fblob\u002Fmain\u002Fplugins\u002Fflow\u002Fskills\u002Ftesting-rules\u002FSKILL.md",[],[],{"skillNo":70,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"publishedAt":19},"01M72CJ6N80HW2WRS8BV4QSS4V","security-rules","Apply application and AI security checks to code reviews, implementation and threat-driven validation.","https:\u002F\u002Fgithub.com\u002F1aGh\u002Fmaude\u002Fblob\u002Fmain\u002Fplugins\u002Fflow\u002Fskills\u002Fsecurity-rules\u002FSKILL.md",[],[],{"skillNo":77,"name":78,"nameZh":6,"description":79,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":80,"categories":81,"keywords":82,"publishedAt":19},"01M72CJ4C41Q4E1WDTFMKYFX3Q","scenario","Author or run cross-platform UI scenarios with selectors, screenshots, platform coverage and result reports.","https:\u002F\u002Fgithub.com\u002F1aGh\u002Fmaude\u002Fblob\u002Fmain\u002Fplugins\u002Fflow\u002Fskills\u002Fscenario\u002FSKILL.md",[],[],{"skillNo":84,"name":85,"nameZh":6,"description":86,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":87,"categories":88,"keywords":89,"publishedAt":19},"01M72CJ1QYX752BM7G12250YJR","question-protocol","Batch necessary questions from workflow participants, collect user answers and route them to the requester.","https:\u002F\u002Fgithub.com\u002F1aGh\u002Fmaude\u002Fblob\u002Fmain\u002Fplugins\u002Fflow\u002Fskills\u002Fquestion-protocol\u002FSKILL.md",[],[]]