L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Productivité · SKILL.md

clip-workflows

clip-workflows

Use FunClip's ASR, text/speaker clipping, subtitle generation, model selection, and Gradio/CLI launch flows safely while routing provider and release work elsewhere.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: clip-workflows description: "Use FunClip's ASR, text/speaker clipping, subtitle generation, model selection, and Gradio/CLI launch flows safely while routing provider and release work elsewhere." disable-model-invocation: true metadata: disco-role: operating license: MIT

clip-workflows

Use this sub-skill when the task is about launching FunClip, choosing the ASR backend, running recognition then clipping, generating subtitles, or diagnosing text/speaker matching problems.

What this covers

  • Gradio launch and deployment flags: --port, --listen, --share.
  • CLI stage 1/2 recognition and clipping workflows.
  • VideoClipper core methods and state handling.
  • Subtitle/time helpers, offsets, repeated text matching, and speaker diarization.
  • Model selection for Paraformer, Fun-ASR-Nano, SenseVoice, and English Paraformer mode.
  • Output directory behavior and subtitle file generation.

Start here

  1. Read references/workflows.md for Gradio, CLI, and programmatic recipes.
  2. Read references/cli-reference.md for exact flags and stage 1/2 command patterns.
  3. Read references/api-reference.md for create_asr_model, build_launch_kwargs, VideoClipper, runner, and the subtitle helpers.
  4. Read references/troubleshooting.md when clipping returns no matches, subtitles fail, or launch options do not behave as expected.
  5. Run scripts/clip_cli_smoke.py with --repo-root to exercise the no-network matching path from any cwd.

Route elsewhere

  • API-key, provider, prompt, and LLM routing details belong in sibling ../llm-providers/.
  • Release archives, checksum packaging, and publication steps belong in sibling ../release-packaging/.
  • Do not use this sub-skill for live model download validation or media encoding verification unless the current run explicitly executes those paths.

Operational notes

  • Keep user-facing examples self-contained: use user-provided media paths, not repository fixture paths.
  • Treat the bundled smoke script as the deterministic check for text matching and no-match behavior; it avoids model downloads and real media files.
  • When a user needs clip-by-text, clip-by-speaker, subtitle overlay, or launch help, stay in this sub-skill instead of reopening the source repository.
Skills enregistrées