Music To Video

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). 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 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 gave you audio — a music file, or a video to pull the audio from — use it. If not, generate one: choose the mood from the user's description (e.g. "driving synthwave", "trap beat", "upbeat corporate") and produce a track via /hyperframes-media (references/bgm.md — HeyGen retrieval when credentialed, else local Lyria / MusicGen; ElevenLabs or another generator also works). Before generating, run npx hyperframes auth status and relay its output verbatim (don't paraphrase or rewrite it) — it shows whether BGM comes from HeyGen or local MusicGen and, if not signed in, how to sign in. If not signed in, STOP and wait for the user to choose — sign in, or continue offline with local MusicGen — before generating the track; don't write keys into a per-repo .env. (In autonomous mode, note the status and continue offline.) See /hyperframes-media → Preflight for the canonical guidance. Either way the track lands at assets/bgm.mp3. Stage any user-supplied images or videos so frames can weave them in on the beat grid; otherwise typography carries the whole video.

Initialize only if hyperframes.json is missing. Name <project> from the brief in kebab-case, such as midnight-drive-loop — never a timestamp.

npx hyperframes init "videos/<project>" --non-interactive --skip-skills --example=blank
mkdir -p "$PROJECT_DIR/assets" "$PROJECT_DIR/renders"
cp "<user-music>" "$PROJECT_DIR/assets/bgm.mp3"   # extract from a video first if needed
# only if the user gave you images/videos:
node <SKILL_DIR>/scripts/stage-assets.mjs --from <dir> --hyperframes "$PROJECT_DIR" --into public

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:

python3 -m pip install librosa numpy soundfile

python3 <SKILL_DIR>/scripts/analyze-beatgrid.py "$PROJECT_DIR/assets/bgm.mp3" \
  -o "$PROJECT_DIR/audiomap.json" --print

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.