Skip to content

Local video editor for coding agents: 42 FFmpeg tools with a machine-readable contract and an MCP server, for Claude Code, Cursor and Codex. Cut, captions, reframe, loudness, platform exports. No cloud, no API keys.

1.5 KB
Raw Blame
{
 "iteration": 2,
 "date": "2026-09-04",
 "runs": 24,
 "prompts": 24,
 "variants": [
  "with_skill (0.8.1 SKILL.md + fixes below)"
 ],
 "prompt_mix": {
  "english_edit": 12,
  "japanese_edit": 8,
  "must_refuse": 4
 },
 "routing_accuracy": {
  "with_skill": 1.0
 },
 "mean_commands": {
  "with_skill": 6.5
 },
 "refusal_honesty": "4/4 (face blur, TTS narration, URL download, speech translation: all declined honestly, nearest alternative offered)",
 "japanese_report_for_japanese_prompt": "9/9",
 "report_format_followed": "20/24 (the 4 misses are refusal / probe-only answers where the Done/Look template does not apply)",
 "visual_check_when_picture_changed": "8/8",
 "findings": [
  "caption.py --text wrote the generated SRT next to the SOURCE file when -o pointed elsewhere; agents then had to explain a stray file. Now written next to the output.",
  "probe.py did not accept --json / the common flags although SKILL.md says every script does; agents retried. Now accepts them.",
  "audio.py --fade-out was applied to the music bed AND the whole mix (double fade on the last seconds). Split into --music-fade-out (bed only) and --fade-out (whole mix).",
  "Japanese caption prompts: agents did not know how to find a CJK font. SKILL.md now shows `fc-list :lang=ja file`.",
  "Not done in this iteration: 3 repeats per prompt and grading by an independent model. Grades are regex-based (evals/grade_runs_24.py) plus manual reading of every run.md."
 ],
 "grader": "evals/grade_runs_24.py",
 "prompts_file": "evals/agent_prompts_24.json"
}