FunkePublishing/skill-promo-video

promo-video

Assemble short promo videos from raw clips. Portrait, square, or landscape. Optional Blender-rendered intro/outro, optional AI-generated music.

The skill is agnostic of the host agent: it MUST work identically across every runtime that implements the Agent Skills standard (https://agentskills.io) — Claude, Claude Code, OpenAI Codex, Gemini CLI, Cursor, GitHub Copilot, Cline, OpenCode, OpenHands, Goose, Kiro, Junie, and others. Never rely on host-specific tools (no AskUserQuestion, no slash commands, no MCP unless explicitly available). Conduct all user interaction as normal chat messages.

When to use

• "Cut these clips together into a 30-second reel."
• "Build me a hackathon promo video in portrait."
• "I have a folder of phone videos — make a 1080x1920 social clip with music."
• "Generate a promo with logo intro + heart outro + background track."
• The user references a directory of .mov/.mp4 files and wants one deliverable.

Prerequisites

• ffmpeg on PATH, built with libx264, aac, and prores_ks (the last one is only needed for alpha-channel intro/outro MOVs).
• Python 3.9+.
• Install Python deps once: pip install -r requirements.txt from inside the skill directory.
• Optional: a .env at the project root containing SUNO_API_KEY to enable music generation. The skill auto-detects whether the key is set and only offers music generation if it is.
• Optional: Blender MCP available in the host agent to produce intro/outro renders. See references/blender_mcp.md.

Default flow (dialog-based)

Step 0 — Detect whether clips have changed since last run

Before asking any questions, check if the user is resuming from a previous session. If gallery/.clips_snapshot.json exists (or if the user points you at a clips folder that was used before), run:

python scripts/clip_diff.py <clips_dir>

• Exit code 0 — clips folder matches the last snapshot. Continue.
• Exit code 1 — drift. The report lists new / removed / modified clips. Show it to the user, then offer to re-run scripts/gallery.py so the gallery picks up the new clips. Existing selections for clips that are still there survive untouched.
• Exit code 2 — no snapshot yet. Normal first-run, continue.

Do NOT skip this step when resuming an existing project — silently working against a stale gallery / selection.json is the classic "where did my new clips go?" bug.

Questions

At the start, ask the user the questions below, in order. Accept natural answers; do not present a multiple-choice UI unless the host supports it. For every question, also accept "you decide" and apply the sensible default.

1. Target format. 1080x1920 portrait (default), 1920x1080 landscape, or 1080x1080 square. Custom sizes allowed.
2. Clips source. Absolute or relative path to a directory containing the raw clips. .mov, .mp4, .m4v are supported.
3. Target length in seconds. (Typical: 20–45 s.)
4. Music. One of: (a) existing audio file, (b) generate via Suno (only offer if SUNO_API_KEY is set), (c) no music.
5. Intro. One of: none, static image (path), Blender MCP generated.
6. Outro. Same choices as intro.
7. Director mode. Yes or no. If yes, follow references/director_mode.md instead of continuing linearly here.

Then handle clip selection. Offer two paths:

A) Visual gallery (recommended for >10 clips or when the user wants to see frames before picking cues). Run scripts/gallery.py <clips_dir> --resolution WxH --fps 30 --target-length N — it generates gallery/gallery.html with thumbnails, IN/OUT sliders, speed slider, and range preview. Tell the user to open it in a browser, pick clips, set cues, then click "Export selection.json". The downloaded file is ready for build.py.

B) Inline dialog (fine for small batches or when no browser is available). Walk through each clip yourself:

• cue_in and cue_out in seconds (trim range)
• speed multiplier (1.0 = original)
• fixed: true if timing/speed is hand-curated, false for free fillers (default: 0.6 s from the middle, speed 1.0)

Offer "Just give me defaults" — pick every clip alphabetically, free defaults, up to the target length.

Either way, the result is a selection.json matching templates/selection.schema.json. Validate before proceeding.

Build

Once selection.json is valid, run:

python <skill>/scripts/build.py <selection.json> --out output/<name>.mp4

build.py writes intermediate artefacts under build/ (relative to the current working directory): build/normalized/, build/trimmed/, build/montage.mp4, build/composed.mp4. Delete build/ to force a full re-render.

Every submodule under scripts/ is also runnable standalone — useful for iterating on one step without rebuilding the whole pipeline. Each takes an --help argument.

Director mode (optional)

Opt-in. See references/director_mode.md for the role prompts. The skill spawns four sequential roles — Narrator → Cutter → Composer → Supervisor — that share selection.json as state. Each role reads the current JSON and writes its updates back. Use this when the user asks for "a more guided process" or "walk me through it like a producer".

Music generation (optional)

If the user chose "generate via Suno", delegate the music step to the Song-Builder sub-agent defined in references/song_builder_agent.md. That sub-agent:

1. Asks the user for genre, mood, BPM range, section structure, instrumental on/off, and takes video length as duration.
2. Formulates a Suno prompt + style tags.
3. Calls scripts/music.generate_song(spec) — supply takes > 1 if the user wants options to compare.
4. Trims and fades the chosen take to the video length.
5. Writes the final mp3 path into selection.json as audio.

If the host provides a sub-agent mechanism (e.g. Claude Code's Agent tool), use it. Otherwise, play the role inline in the same thread.

Blender MCP integration (optional)

Only assistance — this skill does not bundle a Blender scene.

If the user wants a Blender-rendered intro or outro, delegate to the Blender-Builder sub-agent defined in references/blender_builder_agent.md. It owns the full flow: preflight (check that Blender + MCP are actually reachable in this host), design interview (visual direction, colors, duration), scene build, iterative preview, render to ProRes 4444 alpha MOV, and hand-off of the MOV path into selection.json.

If the host provides a sub-agent mechanism (e.g. Claude Code's Agent tool), spawn it. Otherwise play the role inline. The main dialog flow should invoke the Blender-Builder whenever the user chose "Blender MCP generated" for intro or outro in step 5 of the questionnaire.

Setup documentation (install Blender, install the MCP server, register it with the host) lives in references/blender_mcp.md. The Blender-Builder sub-agent pulls from that doc when the user needs to configure the stack for the first time.

Error recovery

• If a clip fails normalization (codec / resolution edge case), build.py exits with a non-zero code and a clear stderr message pointing at the offending file. Ask the user whether to skip the clip or investigate.
• If Suno returns a 4xx or times out, offer the user to retry, switch to "existing file", or proceed without music.
• Never silently drop clips. Never overwrite an existing output file without confirmation unless --force is passed.

Style notes for this skill

• Never assume host-specific tools. All dialog is plain chat.
• Show your work. When trimming, tell the user which cue points you settled on before building.
• Fail loud. If ffmpeg is missing, say so and show the install command for the user's platform.