music-to-video

First, keep this skill fresh — confirm with the user before running: npx hyperframes skills update music-to-video. A fast no-op when everything is current; otherwise it refreshes this skill plus the core domain skills it depends on before you rely on them.

music-to-video — one music-grounded, beat-synced video workflow

Use this skill to turn a music track into a beat-synced HyperFrames video. You analyze the track once, lay out the frames, fill in a per-frame plan, and build each frame as a composition. The input is a music track plus optional user images or videos — there is no narration and no website capture. Typography and templates are the floor (a complete video needs zero assets); any media the user supplies is cut in on the same beat grid.

You are the orchestrator. Work in videos/<project>/. Run the steps in order and pass each Gate before moving on. Two steps need the user: Step 3 (plan approval) and Step 6 (render approval) — both are checkpoint gates per ../hyperframes-core/references/brief-contract.md (read it before Step 0): in autonomous mode, post the summary as a heads-up and proceed instead of waiting. Do every step yourself except Step 4, where you dispatch one sub-agent per frame. Keep design and motion rules out of this file — they live in references/ and the frame-worker sub-agent.

SKILL_DIR = this skill directory. PROJECT_DIR = videos/<project-name>/.

Workflow: Step 0 setup → hyperframes.json + assets/bgm.mp3; Step 1 analyze → audiomap.json; Step 2 skeleton → STORYBOARD.md (frames, groups TBD); Step 3 plan → complete STORYBOARD.md + frame.md; Step 4 build → compositions/frames/NN-*.html; Step 5 assemble → index.html; Step 6 render → renders/video.mp4.

Two ideas that shape everything

• One analyzer, and you trust it. analyze-beatgrid.py is the only beat analyzer — never re-measure beats with another tool or by ear. Its energy / density / rolls / onsets / silences are always reliable. Its bpm and beats_sec are reliable only when the music is genuinely rhythmic; on calm music the grid is a metronome the tracker imposed, so pace by phrases and energy instead and never hard-cut to it. Deciding which case you're in is each frame's pacing (Step 2).
• One frame = one file; groups live inside. Step 2 cuts the track into frames, and each frame becomes one composition file compositions/frames/NN-<frame_id>.html, built by one frame-worker. A frame can subdivide into groups (each a template or a motion-primitives combo). Extra density goes inside a group, so frame count tracks distinct treatments, not beats — a fast track does not blow up the number of sub-agents.

Step 0: Setup, BGM, and inputs

Goal: Establish the music source, create the HyperFrames project, and note any user-supplied media.

The brief starts at the intent layer. Opening rule, in order: (1) BRIEF.md exists → read it and ask nothing it answers — its flow/storyboard derive the mode (brief contract § 1). (2) No BRIEF.md but the project exists → resume from what's on disk; never re-interrogate. (3) A fresh creation request that arrived here directly → read /hyperframes and run its intent layer (references/intent-interview.md): it confirms this route's must-haves (the music source, destination → aspect — ../hyperframes/references/routes/music-to-video.md) and announces what stays deferred — brand and genre are chosen at Step 3 by design. Write BRIEF.md immediately after init (never before — init refuses a non-empty directory) and record the preference-backed answers (brief-format.md). Edit requests skip all of this.

The music is the spine — establish one track before anything else. This skill is tuned for fast, high-energy BGM: a strong beat grid drives the cuts (calm tracks work, but pace by phrase rather than beat). If the user supplied audio — a music file, or a video to pull audio from — use it. Otherwise choose the mood from the request and generate a track through /media-use (references/bgm.md). Before the first authenticated provider action, run npx hyperframes auth status and relay its output verbatim. If signed out, apply one branch:

• Collaborative: wait for sign-in or an explicit choice to continue offline with the local provider.
• Autonomous: state the status and continue through the available local provider.

If no offline provider can satisfy the required music capability, surface the blocker. Never write keys into a per-repo .env. Auth ownership and offline fallbacks live in /media-use references/setup-providers.md § Providers. The resulting track lands at assets/bgm.mp3. Stage supplied images or videos so frames can use them on the beat grid; otherwise typography carries the video.

Lyric videos: for lyrics synced to the vocals, get word/line timing by transcribing the track via /media-use, or ask the user for the lyrics text and place lines on the beat grid.

Initialize only if hyperframes.json is missing. Name <project> from the brief in kebab-case, such as midnight-drive-loop — never a timestamp. init checks the installed skills against the latest on GitHub and updates the global set if any are out of date.

npx hyperframes init "videos/<project>" --non-interactive --example=blank --skill=music-to-video
mkdir -p "$PROJECT_DIR/assets" "

The brand (font + palette) is chosen at Step 3, not here. Don't pick a genre or a track type up front — assets are just an optional ingredient, and the genre emerges from the per-frame choices.

Gate: hyperframes.json + assets/bgm.mp3 exist; aspect / length / fps and (if any) the asset inventory are noted.

Step 1: Analyze the music

Goal: Produce the one canonical timing analysis the whole video is built on.

analyze-beatgrid.py is the only beat analyzer — never re-measure beats with another tool or by ear. It reads the track once and writes audiomap.json: energy phases (level / density / feel), onsets + onset_rate, rolls, silences, hard_stops, key_moments, phrases, tempo / grid, and audio.duration_sec. It's deterministic — the same file always gives the same map. Most fields are reliable on any music; bpm and beats_sec are reliable only when the music is genuinely rhythmic, and judging that is the call you make at Step 2.

Prerequisites: Python 3 with librosa, numpy, and soundfile available. If import fails, install them into the active Python environment before running the analyzer:

Gate: audiomap.json exists; audio.duration_sec is known.

Step 2: Frame skeleton (structure only)

Goal: Read the music and lay out the frames — the skeleton of STORYBOARD.md.

Read references/frame-skeleton.md. Turn audiomap.json into the skeleton of STORYBOARD.md yourself — there is no intermediate JSON. Cut the track into frames at real musical changes (hard_stops, SURGE / DROP key_moments, the edges of a roll, a stretch with no onsets, a big energy jump), snapping every boundary to an audiomap anchor. For each frame set span_sec, pacing (the verdict from Step 1's trust call — beat_cut when the grid is real, phrase_flow when it's a metronome imposed on calm music), mood, and a one-line feel (the plain music situation Step 3 matches a template against). Only classify and lay out here: leave every frame's ### Groups as TBD (Step 3) and the frontmatter style blank — no templates, copy, color, or fonts. Expect ~1–6 frames.

Gate: frames tile the track (first at 0, last at duration_s); each carries span_sec + pacing + mood + feel; every ### Groups is TBD; no content anywhere.

Step 3: Fill the plan (user-gated)

Goal: Turn the skeleton into an approved, complete STORYBOARD.md.

Read references/planning.md, storyboard-format.md, template-catalog.md, motion-primitive-catalog.md, and montage.md (only if the user supplied assets). Editing the same file in place, do two things:

1. Pick the brand. Choose one preset from ../hyperframes-creative/frame-presets/ using the table in ../hyperframes-creative/references/design-spec.md (match the track's mood; only its fonts and colors matter — templates own composition). Copy it into frame.md unmodified and fill the frontmatter style (font + a ≤4–6 swatch palette) from it.
2. Fill every frame. Decide its groups and give each a treatment: a matched template from the catalog (with bound params and real audiomap anchors), a free-compose from the primitive catalog, or an asset treatment that obeys pacing. Before you free-compose a named look, search the live catalog for it: for every look, effect, treatment or transition the user asked for — "CRT scanlines", "glitch", "film grain", "shimmer sweep" — run npx hyperframes catalog --query "<the look, in plain English>" --json and read the top results. template-catalog.md and motion-primitive-catalog.md list only this skill's own local materials; the search ranks the whole hosted registry (~400 blocks and components) and needs nothing installed — no project, no prior add, no account. Free-compose a look only after a search for it came back with nothing that fits. Write the copy. You own WHAT (template / primitives + content + anchors); the frame-worker owns HOW — never write millisecond tweens into the storyboard.

Fix every ✗ (hard errors: duration mismatch, frames not tiling the track, a missing src); warnings are best-effort. Then show the user a frame-by-frame summary and iterate until they approve. In autonomous mode this is a checkpoint gate: post the summary as a heads-up and proceed (the validate-plan.mjs pass is a quality gate and still blocks).

Gate: frame.md is a verbatim preset copy; validate-plan.mjs exits 0; the user approved the plan (autonomous: the summary was posted as a heads-up).

Step 4: Build frames from the plan

Goal: Build every frame as a self-contained composition file.

Create compositions/frames/. Read sub-agents/frame-worker.md and ../hyperframes-core/references/subagent-dispatch.md. Dispatch one frame-worker per frame, in parallel where possible (otherwise in waves). Each worker gets exactly one frame and this context:

The worker forks the cited materials, converts every anchor to frame-local seconds (local_t = track_t − span_sec[0]), gates its groups with 0ms cuts, and writes one seek-safe frame file. The worker never runs the hyperframes CLI — those commands operate on the assembled project, which doesn't exist yet, so they'd report on the wrong files. The worker just writes to the contract and stops; you verify after assembly (Step 6). As each worker returns, you can confirm its file landed on disk.

Gate: every frame has its compositions/frames/NN-*.html on disk.

Step 5: Assemble

Goal: Wire the built frames + BGM into the playable index.html.

assemble-index.mjs is deterministic — no subagent, no judgment. It references each frame file at its cumulative data-start, mounts assets/bgm.mp3 on track 11, and hard-cuts frame → frame (frames tile the track with no gaps, so there is no transition injector).

Fix any ✗ it reports — a missing or blank frame file means that worker wrote a partial file; re-dispatch it (Step 4) and re-assemble.

Gate: index.html exists; total duration == audiomap.audio.duration_sec.

Step 6: Verify and render

Goal: Verify the assembled video, get user approval, and render the final MP4.

Run the CLI on the assembled project — that's the correct unit (the per-frame workers couldn't run it). check runs structural lint and the headless-browser runtime, layout, motion, and contrast gate in one pass; --snapshots also emits the review frames.

Inspect at t=0, each frame start, the strongest DROP / SURGE, every hard_stops[].t, and the final frame. On failure, make the cheapest safe fix yourself: edit the offending compositions/frames/NN-*.html. Never change duration or audio timing to hide a sync issue. Once the gates pass, pause for user review, then render only on approval (autonomous mode: ask the one kept question — "preview first, or render?" — then deliver the MP4 with the contact sheet):

Gate: check passed and the snapshots were inspected; the user approved (autonomous: checks passed and the delivery includes the contact sheet); renders/video.mp4 exists with audio, duration == audiomap.audio.duration_sec. The final reply states the MP4 path and duration.

Resume table

Quick Reference

Formats: landscape 1920x1080 by default; portrait 1080x1920; square 1080x1080. Set the canvas once in the storyboard frontmatter (canvas: { w, h, fps }).

Scripts under scripts/: analyze-beatgrid.py (the one analyzer), validate-plan.mjs (plan check), assemble-index.mjs (index assembly), stage-assets.mjs (stage user media), lib/storyboard.mjs (vendored parser). Everything else is the hyperframes CLI.

Directory layout