lalimi/claude-reelsmith

Finish a built reel in CapCut — upload the picture (everything baked in), the voice and music stems, place CapCut library sound effects on every effect and transition, push an editable CapCut Web draft, and let the creator swap music and sounds by ear before exporting. Uses the CapCut MCP connector bundled with this plugin. Use when asked for "sound effects from CapCut", "put it into CapCut", "editable draft", "change the music in CapCut", "капкат".

What is claude-reelsmith?

claude-reelsmith is a Claude Code agent skill that finish a built reel in CapCut — upload the picture (everything baked in), the voice and music stems, place CapCut library sound effects on every effect and transition, push an editable CapCut Web draft, and let the creator swap music and sounds by ear before exporting. Uses the CapCut MCP connector bundled with this plugin. Use when asked for "sound effects from CapCut", "put it into CapCut", "editable draft", "change the music in CapCut", "капкат".

Works with✓Claude Code✓Codex CLI~Cursor

npx skills add https://github.com/lalimi/claude-reelsmith/tree/HEAD/skills/reel-capcut

Documentation

Reel → CapCut draft

Needs: a built reel (reelsmith build <dir>), the CapCut MCP server (capcut, bundled in this plugin's .mcp.json; sign in once via /mcp), and CapCut's own plugin bundle on disk for the uploader and draft builder (see $(reelsmith root)/docs/capcut.md, env CAPCUT_PLUGIN_DIR, default ~/capcut-plugin).

Steps

Add "sfx" to reel.json — one entry per on-screen event:
```
"sfx": [{"name": "whoosh", "t": 6.9, "vol": 0.5}, {"name": "click", "t": 30.62, "vol": 0.5},
        {"name": "typing", "t": 2.5, "vol": 0.35, "dur": 1.4}, {"name": "my_sounds/rip.mp3", "t": 15.36}]
```
Library names: whoosh, whoosh-short, pop, typing, click, sparkle, coin, chime, impact, shutter, ding (ids in engine/capcut.py). Anything else = a local audio file you own, uploaded with the stems.
reelsmith capcut stems <dir> → out/capcut/{video.mp4, voice.wav, music.wav} + files.json. Voice and music share one gain (sum = −14 LUFS) and a limiter; music is already ducked under the voice.
Upload: call capcut_asset_upload_credential with version: 2 and every file's exact file_name + size_bytes (re-check sizes right before; a mismatch fails with upload_files_mismatch). Then printf '%s' <ticket> | node "${CAPCUT_PLUGIN_DIR:-$HOME/capcut-plugin}"/x/plugins/codex/skills/capcut-upload/scripts/capcut-upload.mjs --files "$(cat out/capcut/files.json)" --upload-ticket-stdin --json > out/capcut/upload.json
reelsmith capcut draft <dir> → out/capcut/draft.json (validated).
start_stdraft_session → push_stdraft_snapshot with base_package_id: "" and the draft JSON. Give the creator the web_editor_url.
Before pushing again, get_stdraft_snapshot and build on it — the creator may have edited the draft (read-back is lossy; never overwrite their changes blindly). For small fixes (trim the music tail, volume) prefer doing them in the editor through Claude in Chrome.

What to tell the creator

Listen first: CapCut library sounds have been reported to sit on the timeline but play silent. If so, replace them in Audio → Sound effects, or use local files.
Music: mute the placeholder stem and pick a library track by ear; trim it to the video length (a longer track adds a black tail to the export).
Export (top right → Download, 1080p / 30 fps / MP4) saves to their CapCut cloud first: if the cloud space is full, the export fails with "Couldn't save" — free space in their CapCut space or keep the engine's MP4.
The connector's own sound search (search_capcut_material) is a poor title search without previews; the full sound library is only in the CapCut app.