magnus919_agent-skills/ffmpeg/evals/evals.json

{
  "schema_version": 1,
  "skill_name": "ffmpeg",
  "evals": [
    {
      "id": "inspect-before-remux",
      "prompt": "I have an MKV with several tracks and need an MP4 without re-encoding. Give me a safe approach.",
      "expected_output": "Explain that this is a remux/stream-copy operation, probe the input first, make stream selection explicit, use no-overwrite behavior during exploration, and probe the output because container compatibility is not guaranteed.",
      "assertions": [
        "distinguishes remuxing from transcoding",
        "uses ffprobe before the operation",
        "includes explicit mapping or explains why mapping is needed",
        "mentions container and stream compatibility",
        "includes post-run output verification"
      ],
      "case_set": "dev"
    },
    {
      "id": "build-aware-filter",
      "prompt": "This command fails with No such filter: drawtext. Replace it with something that will work everywhere.",
      "expected_output": "Do not claim a universal replacement. Explain that filter availability is build-dependent, show how to inspect the installed filter inventory/help, and offer a conditional path: use an available equivalent, install a build containing the filter, or omit the transformation.",
      "assertions": [
        "does not fabricate universal filter availability",
        "checks ffmpeg filter capabilities",
        "treats the error as a build boundary",
        "offers conditional next steps rather than silently substituting",
        "preserves the need to rerun and verify the output"
      ],
      "case_set": "regression"
    },
    {
      "id": "explicit-complex-mapping",
      "prompt": "Overlay a logo image on the first video stream while keeping the first input's audio. Show the important FFmpeg structure.",
      "expected_output": "Use a complex filtergraph with two labeled video inputs, label the overlay output, map that label exactly once, map the selected audio explicitly, and explain that filtering requires decoding and video re-encoding while audio may be copied.",
      "assertions": [
        "uses filter_complex with two inputs",
        "labels and maps the filtered video output",
        "maps the intended audio stream explicitly",
        "explains filter and stream-copy boundaries",
        "does not imply automatic mapping is safe here"
      ],
      "case_set": "dev"
    },
    {
      "id": "timestamp-concat-diagnosis",
      "prompt": "Two clips have matching extensions but concatenation produces a jump and audio drift. What should I inspect and which concat mechanism should I choose?",
      "expected_output": "Treat the issue as a stream-compatibility and timestamp problem. Compare codecs, dimensions, rates, channel layouts, start times, durations, time bases, and packet/frame behavior; distinguish concat demuxer, concat filter, and concat protocol; normalize timestamps and formats when using the filter.",
      "assertions": [
        "does not use file extension as the compatibility test",
        "lists relevant audio/video stream properties",
        "distinguishes concat demuxer from concat filter and protocol",
        "identifies timestamps as a diagnosis surface",
        "recommends probing inputs and output"
      ],
      "case_set": "regression"
    },
    {
      "id": "safe-batch-network",
      "prompt": "Write a shell-loop recipe that converts every file in a directory and sends results to a network endpoint.",
      "expected_output": "Before giving a recipe, set safe boundaries: use arrays or null-delimited discovery, avoid shell concatenation of untrusted filenames, refuse overwrites by default, record command/version/exit status, probe each output, and treat network protocol, timeout, authentication, and destination validation as explicit requirements.",
      "assertions": [
        "addresses shell quoting and untrusted filenames",
        "uses a deliberate overwrite policy",
        "requires per-output verification and logging",
        "calls out network protocol and timeout assumptions",
        "does not expose credentials in command arguments"
      ],
      "case_set": "release"
    },
    {
      "id": "hardware-pipeline-caveat",
      "prompt": "NVENC is enabled. Does that prove hardware acceleration will make this workflow faster on my Mac?",
      "expected_output": "No. Explain that decoding, filtering, memory transfers, and encoding are separate pipeline decisions, that NVIDIA-specific NVENC guidance does not transfer to macOS, and that the exact local build/device/pixel-format path must be inspected and benchmarked end to end.",
      "assertions": [
        "rejects encoder-only speed conclusions",
        "separates decode/filter/encode and memory-transfer decisions",
        "recognizes platform-specific hardware guidance",
        "requires local capability inspection",
        "requires reproducible end-to-end benchmarking"
      ],
      "case_set": "release"
    }
  ]
}