AutoDesign Video

Create the video with the host coding agent. Use the bundled harness for evidence, durable attempts, exact HyperFrames rendering, local Kokoro narration, optional subtitles, media QA, and hash-bound delivery. This Skill does not require the AutoDesign repository, server, or an API judge.

Read first

Read source-grounding.md, output-contract.md, and review-rubric.md completely. Resolve SKILL_ROOT to this installed Skill folder. Put the run in a user-selected output directory outside the Skill. Never write output, caches, models, voices, browsers, virtual environments, or node_modules inside the installed Skill.

Execute

Choose a supported Python launcher first: try python3 --version, then python --version, and on Windows py -3 --version; use the first command that reports Python 3.10–3.12. Run that exact launcher (including both tokens in py -3) with $SKILL_ROOT/scripts/video_harness.py --help and $SKILL_ROOT/scripts/setup_video.py --help for exact flags.

1. Run doctor. If it reports a missing runtime, run setup. Setup requires Node 22+, npm, ffmpeg, ffprobe, Python 3.10–3.12, and the Poppler PDF-ingest commands pdftotext, pdfinfo, pdftoppm, and pdfimages. Doctor reports every missing command before a paper run. Setup installs exact hyperframes@0.7.86, its rendering browser, kokoro-onnx==0.5.0, and soundfile==0.14.0 in an atomic versioned user cache. It prefetches and SHA-256 verifies the complete platform-aware Python lock, exact Kokoro model, voice blobs, and installed HyperFrames browser. Doctor launches that exact browser and rejects a writable or tampered Python package cache. Never substitute a global or newer HyperFrames version. Unsupported platforms fail closed. Setup does not modify the Skill or global packages; setup_video.py remove deletes only this version's locked cache and is safe to repeat.
2. Run init RUN, then evidence RUN PAPER. Pass user-provided content images with --asset and style references with --reference. Read evidence/evidence.jsonl, the rendered paper pages, and evidence/source_visuals.json. A fresh host VLM or subagent must inspect PDF visual candidates against their captions before bind-visuals. Uncertain visuals stay unused. References are style-only; never copy their content.
3. Author and save a complete input plan, then run plan RUN input-plan.json. Default to 12 scenes and 360 seconds. Honor explicit user overrides only within 10–14 scenes and 300–600 seconds. Build one spoken research argument: question, gap, contribution, method, strongest evidence, analysis, limitations, implications, and closing. Every scene needs English narration, contiguous timing, evidence IDs, and any approved visual IDs. Give each scene an exact title_claim_id, narration_claim_id, non-empty visible_claim_ids, and an allowed visual_role. From this point on, use the canonical bytes at RUN/plan.json; validate and deliver reject any independently reserialized or edited plan.
4. Run begin-attempt RUN. Author a local-only editable HyperFrames project in a work folder for that attempt, not in its reserved artifact/ directory. Follow the exact DOM and security contract in output-contract.md. Use native HTML/CSS/SVG text, real paper figures, deterministic seekable timelines, one composition root, literal class="clip" scenes, literal narration audio, and a working subtitle toggle. Give the composition root and every content scene an opaque pure-white rendered canvas so white-background paper figures integrate cleanly. Keep these primary roots at full effective opacity and free of filters, masks, blending, and clipping; animate an inner content wrapper instead. Each scene's actual source-bound image IDs must equal its canonical visual_ids exactly. Do not add inline on* handlers, remote resources, or runtime network behavior.
5. Write a non-empty claims JSON list. Every planned title and narration must equal the text of its named claim exactly; visible numbers must occur in one of that scene's visible_claim_ids. Every claim cites real evidence IDs. Run validate PROJECT RUN/plan.json --run RUN --claims claims.json. This is structural validation; it intentionally does not require narration audio yet and never creates a placeholder. It reuses the shared eligible-role/reuse-limit visual planner and rejects ambiguous HTML, unsafe CSS paths, navigation, and runtime network behavior. Fix every reported source, timing, local-asset, or HTML error.
6. Run deliver PROJECT RUN/plan.json --run RUN --attempt ID --claims claims.json. The harness enforces this order: structural validation; per-scene HyperFrames/Kokoro TTS and timed WAV mix; transcript/SRT/VTT and metadata; strict offline Chromium interaction of every visible enabled native or ARIA control, with identity/result evidence and at least 500 ms of timer, request, navigation, and popup quiescence after every operation. Subtitle checks include ancestor opacity and clipped viewport intersection. The same browser audits the composition root and every scene at initial load, after each operated control, and in the final state, rejecting transparent, tinted, dark, gradient, image, translucent, filtered, masked, blended, or clipped primary canvases. For an animated project it also seeks each scene midpoint through the existing player or unique composition timeline; a missing seek API fails closed because active scenes cannot be verified. Use data-no-timeline only when no player or timeline registry exists and the composition/scene roots have no active CSS animation, transition, or Web Animation. Local motion belongs on descendant wrappers and panels; full real HyperFrames lint; strict real HyperFrames render; subtitle mux; exact ffprobe; representative frames and contact sheet. ffmpeg may mix audio, mux subtitles, and extract frames; it must never replace HyperFrames as the final renderer. A stale or invalid MP4 is deleted and cannot pass. Publishing uses an exact generated allowlist through a sibling staging directory and atomic promotion. The promotion moves the pre-created empty destination aside for Windows compatibility and recovers interrupted stage/backup transactions; copy failures are retryable and never expose a partial live artifact. Hidden files, .env, debug files, and unreferenced assets fail instead of leaking into the artifact. Before any generated media is written, every existing project path is checked for symlink, hard-link, and containment escapes. Published delivery reports use only stable project-relative paths.
7. Run review-context RUN ID. Give the exact MP4, narration WAV, contact sheet, and all six individual frames to a fresh vision- and audio-capable host or subagent that did not author the project. The returned context exposes hash-bound readable source text, evidence JSONL, and the source map; provide them to the reviewer so it can verify semantics rather than guess from frames. The reviewer must inspect every frame, listen to the narration, and return the exact hash-bound schema in review-rubric.md. Record it with record-review; never infer visual quality from HTML or self-certify. A passing verdict requires every rubric dimension to score at least 4/5; there is no averaging exception.
8. If an authoring gate fails, start the next bounded attempt and repair the localized findings. If setup, TTS execution, ffmpeg, ffprobe, or another deterministic runtime stage fails, repair the runtime and resume the same attempt; do not spend an authoring attempt on it. A lint or render failure reruns the exact runtime/browser doctor before it is classified, so browser infrastructure failures remain in the same attempt. On a passing review, run finalize RUN ID and deliver the complete final/ directory together with its run-level QA record.

Run resume RUN before every continuation after interruption. It verifies the installed Skill snapshot, source, attempt artifacts, previews, reviews, and final delivery hashes. A scaffold, silent video, burned-in-only subtitle, plain-ffmpeg slideshow, stale render, failed lint/probe, partial review, or fallback file is not success.

Design standard

Make a conference film, not a narrated slide dump. Use editorial scene composition, restrained academic typography, purposeful motion, direct visual continuity, readable source figures, and evidence-led narration. Keep captions optional and default them off in the authored player. Use opaque pure white #fff with no background image for the composition root and every rendered content scene. Do not fade, filter, mask, blend, or clip these roots; apply motion to their inner content. Player chrome, subtitles, overlays, and restrained light local panels are not primary canvases and may use purposeful contrast. Avoid repeated cards, gratuitous gradients, decorative stock media, fake logos, tiny figures, dense paper paragraphs, constant motion, and generic AI-marketing language.