media-studio

The only public video/audio/animation plugin. Infer form from the ask (product demo, briefing, explainer, talking-head, and the rest) and run that workflow. Live-product screencasts are this pack (Playwright WebM + Recast), not a second plugin.

Rapid default: a sentence or a source path is enough. Infer kind and delivery. Still run curation, gap-closure, writer, storyboard, visuals, director, Azure-first generate, Remotion/FFmpeg/Recast, craft critic, QA. Rapid means infer-and-run, not skip craft. Ask the user only when blocked.

Live-product screencasts use Playwright WebM + Recast in this pack (kit/screencast/record-job.mjs, scripts/validate-product-picture.mjs).

Viewer-facing jobs apply skills/video/references/story.md. The screenplay carries story role, WIIFM, and tension; the storyboard carries shot plan, archetype, sound and one eye path per screen beat. Score < 85 from the story-experience review (review) does not ship. intent: draft skips generated plates and the critic gate. Product-screencast craft uses the product-picture contract. Before capture, scripts/validate-narration.mjs holds the locked script to the no-fluff rule (every line earns its seconds; no greeting, meta, recap, process, version or synthetic-data talk; hook and promise inside five seconds; synthetic disclosure as a caption or end card; fluff per minute and a words-cut report) and scripts/validate-demo-scenario.mjs requires a source-hashed synthetic scenario with one persona, problem, hero, visible outcome, coherent case/target continuity, and a CTA; record-job.mjs runs scripts/validate-capture-preflight.mjs (scene ids, cue keywords, typing budgets) before the first take. After recording, scripts/validate-pointer-path.mjs fails any take whose cursor reversed, wandered, over-travelled or scrolled past and returned. Every ratio is captured at its own viewport with device emulation (job.captureViewports, record-job.mjs --ratio), never cropped out of a desktop take; scripts/validate-picture-integrity.mjs fails stretched, upscaled, cropped, barred or undeclared-zoom product footage and a vertical cut recorded at a desktop viewport. After every recut, scripts/validate-viewer-release.mjs requires exact-candidate captions, loudness, authored-script pronunciation, visual/title/action-result reviews, the narration check, frame integrity (scripts/validate-frame-integrity.mjs), picture integrity, the pointer-path report, the marketing floor on marketing cuts (scripts/validate-marketing-readiness.mjs, references/marketing-readiness.md) and a stitchable sequence before release. scripts/build-contact-sheet.mjs tiles the cue-sheet stills for the human watch-through.

Technical stories (programForm: technical-story) are product or engineering changes told as 5–8 minute mini documentaries: skills/video/references/story.md holds the beat map and markup rules; kit/screencast/ holds the compositor, annotation overlay, code-reveal and card renderers with their manifest hints. Product-specific capture notes live under skills/capture/references/.

Product picture (skills/video/references/product-picture.md): when the subject is a running product, the encoded film is Playwright WebM + Recast (pointer, click, zoom, visible result), not Ken Burns stills mixed with Remotion cards. scripts/validate-product-picture.mjs fails closed. Host projections must match plugin.json.

Version authority: root plugin.json (2.1.3). Host projections must match. Runtime media stays outside AgentHub (%LOCALAPPDATA%\AgentHub\media-studio and MEDIA_STUDIO_ROOT, default D:\Media-Studio). Private show bibles and customer recordings stay outside this repository.

Canonical Remotion kit: kit/ (scene-archetype library, plus kit/screencast/ compositor tooling). Runtime copy: %LOCALAPPDATA%\AgentHub\media-studio\briefing-kit. Sync with scripts/Sync-MediaStudioKit.ps1.

Marketing ads (15-30 s social and web ads in 16:9, 4:5 and 9:16): kit/marketing/ is the generator (mvg.mjs). A spec of beats, claims with evidence, a voice, an optional music bed and a CTA goes through lint, voice, timing, assets, capture reuse, the MarketingAd Remotion composition, QA (including OpenAI vision review of sampled frames) and delivery; spend is capped per run and cumulatively, and every claim must map to evidence. See kit/marketing/README.md.

Pre-capture screenplay lint: scripts/validate-narration.mjs <jobRoot> then scripts/write-story-review.py <jobRoot> (the sync script copies them to _shared\tools; the story review fails no-fluff without the narration check). It is not a craft pass and does not authorize compose or delivery. Product-screencast still requires validate-product-picture.mjs, validate-demo-scenario.mjs, validate-viewer-release.mjs, and encoded-file review (product-picture.md, RELEASE-EVIDENCE.md).

Vocabulary reuse alone can misclassify a truth-status qualification. A job may declare narrationScopeQualifications: [{sceneId, line, reason, evidence: {path, sha256}}] for an exact sentence explicitly marking proposals as subject to client validation. The in-job JSON evidence must carry kind: "truthful-scope-qualification", verdict: "PASS", assertion: "proposed-not-client-confirmed", the exact sceneId and line, and a substantive reason. This adjudicates only the cumulative vocabulary heuristic; banned language, single-line redundancy and repeated sentences still fail. The report exposes each adjudication and hashes its evidence; release rejects changed evidence or declarations. It does not establish client confirmation or owner approval.

Product-repository boundary

Product repositories are read-only inputs. Media Studio never adds or changes video code, Playwright specs, Remotion apps, dependencies, configuration, seed fixtures, assets, evidence, media, or Git state there. The job config and all production source live in the external job workspace. Product defects become readiness feedback and require a separate engineering task.

Why this exists

Product Demo Studio is a specialized, fail-closed product screencast pipeline (Playwright capture, Recast pointer/click, four-domain review). Agents were routing briefings, training films, explainers, talking-heads, and motion-slide pieces through it because it was the only video pack.

This package is the parent studio. Product Demo Studio remains the product-screencast workflow inside it.

Crew (skills, not a second control plane)

Skill	Owns	Key references
video	Classify the job, write the script, storyboard, direct, run the pipeline	story.md, forms.md, scenes.md, product-picture.md
generate	Azure-first voice, ASR and stills under spend controls; Deepgram as governed fallback; accepted music assets	azure.md, content-controls.md, provider-plugins.md
capture	Playwright WebM + Recast	recording.md, abacare-local.md
compose	Remotion, Recast, FFmpeg, Finish-Media.ps1, delivery, optional Descript finishing	render-lanes.md, descript.md
review	Release gates, word clock, the 0-100 story-experience review (fail below 85)	gates.md, word-clocks.md, story-review-rubric.md, marketing-readiness.md

video is the entry point; the other four run single steps. The series and explanation schemas live in schemas/series/ and schemas/explanation.schema.json. Viewer-facing briefing, training, explainer, and series-episode require Remotion unless the user asked for a draft/basic cut. Do not vendor Remotion rules — resolve video.programmatic-composition to official remotion-dev/skills.

Non-goals

• Do not convert Product Demo Studio into a general filmmaking engine.
• Do not centralize credentials, customer data, or private evidence here.
• Do not start local models, TTS/STT servers or embedders: Local AI is off. Hosted providers only, through the governed adapters.

Governed Sienna TTS: scripts/deepgram_cli.py quotes by default; explicit execution requires intact shared accounting or an independently reviewed owner-authorized bounded epoch. See skills/generate/references/content-controls.md. Offline tests: python -m unittest discover -s scripts -p test_governed_deepgram.py. No provider calls are required for tests.

Cloud render: kit/cloud/cloud-render.mjs runs FFmpeg, Remotion and Recast jobs on a scale-to-zero Azure Container Apps job in the owner's own Azure tenant (upload -> render -> download). Provision with kit/cloud/Provision-CloudRender.ps1; see kit/cloud/README.md. Offline tests: node --test tests/cloud-render.test.mjs.

Runware training stills: kit/cloud/cloud-ai.mjs illustration provides an explicitly selected, synthetic-only, single-model job with resumable polling, a $1 per-job spend bound, private hash-bound output, a content-hashed output inventory (jobs/runware/<id>/inventory.json) and human review. See kit/cloud/README.md for the job and retention contract. It does not publish an asset, authorize customer data or establish provider entitlement. Offline tests: node --test tests/runware-image.test.mjs.