video-skill（视频技巧）

Video Skill

Use this skill to run video-skill end-to-end or stage-by-stage.

First-time setup (no repo clone required)

Use one of these setup paths:

A) Run from local source repo (recommended while iterating):

cd /path/to/videoskill
uv sync --dev
cp config.example.json config.json

Then run commands with uv run, for example:

uv run video-skill --help

Then run video-skill ... directly from your working directory.

Verify providers before processing:

video-skill config-validate --config config.json
video-skill providers-ping --config config.json --path /v1/models

Standard workflow (recommended)

Run from your working directory where config.json and data paths are valid.

video-skill transcribe --video <video.mp4> --out <name>.whisper.json --config config.json
video-skill transcript-parse --input <name>.whisper.json --out <name>.segments.jsonl
video-skill transcript-chunk --segments <name>.segments.jsonl --out <name>.chunks.jsonl --window-s 120 --overlap-s 15
video-skill steps-extract --segments <name>.segments.jsonl --clips-manifest <clips>.jsonl --chunks <name>.chunks.jsonl --mode ai --config config.json --out <name>.steps.ai.jsonl
video-skill frames-extract --video <video.mp4> --steps <name>.steps.ai.jsonl --out-dir <frames_dir> --manifest-out <name>.frames_manifest.jsonl --sample-count 2
video-skill steps-enrich --steps <name>.steps.ai.jsonl --frames-manifest <name>.frames_manifest.jsonl --out <name>.steps.enriched.ai.jsonl --mode ai --config config.json
video-skill markdown-render --steps <name>.steps.enriched.ai.jsonl --out <name>.md --title "<Title>"

Modes

• --mode heuristic: deterministic, no model calls
• --mode ai-direct: VLM-centric enrichment
• --mode ai: reasoning + VLM orchestration (default for quality)

Prefer --mode ai unless user asks for debugging or reduced model usage.

Reliability and diagnostics

steps-enrich emits:

• per-step progress logs
• summary metrics: parse_errors, transient_recovered, unresolved_final
• detailed *.errors.jsonl when any errors occur

If runs fail unexpectedly:

1. re-run providers-ping
2. inspect *.errors.jsonl by stage (sampling_plan, vlm_judge, vlm_select_frames, vlm_signal_pass, reasoning_finalize)
3. verify endpoint DNS/host reachability

Validation gate before claiming success

Always run:

video-skill --help

Use make verify only when working from the source repo.