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):
CODEBLOCK0
Then run commands with uv run, for example:
CODEBLOCK1
Then run video-skill ... directly from your working directory.
Verify providers before processing:
CODEBLOCK2
Standard workflow (recommended)
Run from your working directory where config.json and data paths are valid.
CODEBLOCK3
Modes
- -
--mode heuristic: deterministic, no model calls - INLINECODE5 : VLM-centric enrichment
- INLINECODE6 : reasoning + VLM orchestration (default for quality)
Prefer --mode ai unless user asks for debugging or reduced model usage.
Reliability and diagnostics
INLINECODE8 emits:
- - per-step progress logs
- summary metrics:
parse_errors, transient_recovered, INLINECODE11 - detailed
*.errors.jsonl when any errors occur
If runs fail unexpectedly:
- 1. re-run INLINECODE13
- inspect
*.errors.jsonl by stage (sampling_plan, vlm_judge, vlm_select_frames, vlm_signal_pass, reasoning_finalize) - verify endpoint DNS/host reachability
Validation gate before claiming success
Always run:
CODEBLOCK4
Use make verify only when working from the source repo.
视频技能
使用此技能可端到端或分阶段运行video-skill。
首次设置(无需克隆仓库)
选择以下任一设置路径:
A) 从本地源代码仓库运行(迭代时推荐):
bash
cd /path/to/videoskill
uv sync --dev
cp config.example.json config.json
然后使用uv run运行命令,例如:
bash
uv run video-skill --help
之后可直接从工作目录运行video-skill ...。
处理前验证提供商:
bash
video-skill config-validate --config config.json
video-skill providers-ping --config config.json --path /v1/models
标准工作流程(推荐)
在config.json和数据路径有效的工作目录中运行。
bash
video-skill transcribe --video --out .whisper.json --config config.json
video-skill transcript-parse --input .whisper.json --out .segments.jsonl
video-skill transcript-chunk --segments .segments.jsonl --out .chunks.jsonl --window-s 120 --overlap-s 15
video-skill steps-extract --segments .segments.jsonl --clips-manifest .jsonl --chunks .chunks.jsonl --mode ai --config config.json --out .steps.ai.jsonl
video-skill frames-extract --video --steps .steps.ai.jsonl --out-dir dir> --manifest-out .framesmanifest.jsonl --sample-count 2
video-skill steps-enrich --steps .steps.ai.jsonl --frames-manifest .frames_manifest.jsonl --out .steps.enriched.ai.jsonl --mode ai --config config.json
video-skill markdown-render --steps .steps.enriched.ai.jsonl --out .md --title
模式
- - --mode heuristic:确定性,无模型调用
- --mode ai-direct:以VLM为中心的增强
- --mode ai:推理 + VLM编排(默认质量模式)
除非用户要求调试或减少模型使用,否则优先使用--mode ai。
可靠性与诊断
steps-enrich输出:
- - 每步进度日志
- 汇总指标:parseerrors、transientrecovered、unresolved_final
- 发生错误时的详细*.errors.jsonl文件
如果运行意外失败:
- 1. 重新运行providers-ping
- 按阶段检查*.errors.jsonl(samplingplan、vlmjudge、vlmselectframes、vlmsignalpass、reasoning_finalize)
- 验证端点DNS/主机可达性
确认成功前的验证门
始终运行:
bash
video-skill --help
仅在从源代码仓库工作时使用make verify。