Agent Skills: ffmpeg Operations

ffmpeg Operations

Operational expertise for ffmpeg/ffprobe: the ~30 commands that cover most real work, the footguns that silently ruin output, EDL-driven editing (edit-as-code), and eight scripts that replace the logic an agent would otherwise re-derive every task.

Doctrine: probe first

Never transcode, cut, or filter blind. Every media task starts by probing the input — codec, duration, frame rate (constant or variable?), pixel format, rotation, stream layout. Half of all "ffmpeg did something weird" reports are a property of the input the command never checked.

python skills/ffmpeg-ops/scripts/probe-media.py input.mp4            # human summary
python skills/ffmpeg-ops/scripts/probe-media.py --doctor input.mp4   # TRIAGE: hazards + exact fixes
python skills/ffmpeg-ops/scripts/probe-media.py --json input.mp4 | jq '.data.streams'
python skills/ffmpeg-ops/scripts/probe-media.py --keyframes-near 92.5 input.mp4

--doctor makes the doctrine self-enforcing: VFR, HDR transfer, rotation metadata, interlacing, non-yuv420p delivery, and moov-at-EOF each come back as a finding with the exact fix command, and exit 10 means "fix before processing". The --keyframes-near form answers "can I stream-copy a cut at 92.5s?" — it reports the nearest keyframes so you know whether a copy cut will snap (see Footguns). When a command fails with a cryptic message, decode it: references/error-decoder.md.

Before recommending an encoder, verify the build has it. Installed ffmpeg builds vary wildly (especially hardware encoders — listed ≠ working):

bash skills/ffmpeg-ops/scripts/capability-scan.sh           # full: proof-encodes each hw encoder
bash skills/ffmpeg-ops/scripts/capability-scan.sh --quick   # list-only, no GPU touch

Cookbook

Copy-paste recipes, each with the trap it avoids in its comments: references/cookbook.md. Commands are bash-form (see Windows notes); always pass -y or -n so no agent-run command goes interactive.

| Need | Cookbook section | |---|---| | Web H.264, HEVC, AV1, remux, VFR normalize, FFV1 archive, fit-to-size | Convert and compress | | Copy vs frame-accurate trim, concat demuxer vs filter | Cut and join | | Scale, crop, 9:16 blurred pad, rotate, fps, speed, slow-mo, timelapse | Resize, transform, retime | | Watermark, drawtext timecode, burned vs soft subtitles | Overlay, text, subtitles | | Extract, replace, mix, loudness, trim silence | Audio | | 16 kHz mono STT extraction, piping, silence chunking | Speech-to-text prep | | Thumbnails, contact sheets, GIFs, chapters, dataset frames, sprites | Images, GIFs, frames | | Corruption check, frame hashes, metadata strip, ffprobe one-liners | Diagnostics and validation | | Post-download yt-dlp patterns | yt-dlp interop | | Synthetic test fixtures | Generative/test sources |

Footguns

The table that pays this skill's rent. Each row is a class of silent failure.

| Footgun | The trap | The rule | |---|---|---| | -ss + -c copy | Cut starts seconds early or with frozen/black lead-in (snapped to prior keyframe) | Copy cuts snap. Check probe-media.py --keyframes-near; re-encode when exact | | Output-side -to after input-side -ss | Timestamps reset at the seek point, so -to silently becomes a duration | Keep -ss/-to on the same side of -i (both input-side is fast and absolute) | | Missing -pix_fmt yuv420p | Encode "works" but Safari/QuickTime/TVs show black or refuse to play (defaulted to yuv444p/yuv422p from a high-quality source) | Always set it for delivery H.264/H.265 | | Missing -movflags +faststart | Browser can't start playback until the whole file downloads (moov at EOF) | Always set it for web-served MP4 | | Default stream selection | ffmpeg picks ONE stream per type (highest-res video, most-channels audio) — extra audio tracks and all subs are silently dropped | -map 0 to keep everything, explicit -map otherwise | | -vf + -c:v copy together | Hard error — filters require decoding | Filtering implies re-encode; pick one | | VFR source (phone/Zoom/Loom/screen-rec) | Cut math drifts, concat desyncs, players stutter | Normalize first: -fps_mode cfr -r 30 + re-encode (cookbook) | | -vsync (deprecated) | Old flag, removed direction | Use -fps_mode (cfr/vfr/passthrough) | | scale=W:-1 | Odd height → encoder error with yuv420p | Always -2 | | concat demuxer on mismatched inputs | "Works" then glitches/desyncs at boundaries (codec/timebase mismatch) | Demuxer = identical params only; else concat filter with re-encode | | amix default | Each input's volume halved (normalize defaults on) | amix=...:normalize=0 + explicit volume= | | One-pass loudnorm | Dynamic mode pumps quiet passages; output silently 192 kHz | Two-pass linear via loudnorm-scan.py; add -ar 48000 | | -shortest absent on audio-replace | Output runs as long as the LONGEST input (silence or frozen frame tail) | Add -shortest when muxing separate A/V | | BT.601/709 colour shift | Slightly wrong colours after scaling SD↔HD (matrix guessed from resolution) | Tag explicitly when it matters: see references/color-hdr.md | | drawtext/subtitles path escaping | Filter args re-parse : and \ — Windows paths like C:\x explode inside filter strings | cd to the asset's dir and use bare relative names; or escape as C\:/path | | Interactive overwrite prompt | Agent-run command hangs forever on "File exists. Overwrite? [y/N]" | Always pass -y or -n explicitly | | % in cmd.exe | %06d patterns and %{pts} get mangled by cmd variable expansion | Use PowerShell or bash; in .bat double to %% |

Windows notes

Platform-agnostic commands, but when running on Windows:

• PowerShell quoting is friendlier than bash here: single quotes are fully literal, so -vf 'scale=1280:-2,fps=30' needs no escaping. Double quotes only interpolate $ and backtick — filtergraphs rarely contain either.
• NUL not /dev/null for two-pass logs: -passlogfile defaults are fine, but ffmpeg ... -f null NUL (PowerShell also accepts -f null -, which is portable — prefer it).
• Font paths in drawtext: fontfile='C\:/Windows/Fonts/arial.ttf' — forward slashes, escaped drive colon, inside the filter string.
• Prefer -f null - and relative paths to sidestep both quoting tables at once.

Decision trees

Codec — H.264 (libx264): default; universal playback, fast, good per-bit at -crf 18..23. → H.265 (libx265): same quality ~40% smaller; slower; needs -tag:v hvc1 for Apple; fine for storage/modern devices. → AV1 (libsvtav1): best compression, royalty-free, web-first (YouTube/Netflix path); encode cost highest; playback on older hardware is software-only. → VP9: only when a pipeline demands webm and AV1 is unavailable. → FFV1: archival masters only.

Cut method — Need exact frames OR applying any filter → re-encode (input-side -ss, -crf 18). Cut points happen to sit on keyframes (verify with --keyframes-near) OR a ±2s slop is acceptable → stream copy with -avoid_negative_ts make_zero. Many cuts from one source → EDL workflow below.

CPU vs hardware encode — Hardware (NVENC/QSV/AMF/VideoToolbox) is 5-20× faster but worse quality per bit than libx264/x265 at slow presets. Use hardware for: speed-critical batch work, live/streaming, drafts, "good enough" deliveries (bump bitrate ~30% to compensate). Use CPU for: final masters, size-constrained targets, quality comparisons. Always capability-scan.sh first — listed encoders fail at runtime on driver mismatches. Details: references/hardware-accel.md.

EDL workflow (edit-as-code)

For any multi-cut edit, do not fire ad-hoc trim commands. Write an edit decision list — a JSON file naming every clip, time range, and why — then cut from it. The edit becomes reviewable (rationale is written down), rerunnable (regenerate the output any time), and diffable (versions of the edit are git history).

# 1. Find candidate cut points (silence = clean speech boundaries):
python skills/ffmpeg-ops/scripts/detect-segments.py --silence --json take3.mp4

# 2. Author the EDL (schema: assets/edl-schema.json) with per-scene rationale.

# 3. Dry-run prints every ffmpeg command it would run (default — nothing executes):
python skills/ffmpeg-ops/scripts/cut-from-edl.py edit.json

# 4. Execute: cuts + concat -> final. Re-encodes by default for frame accuracy;
#    --copy for keyframe-aligned EDLs.
python skills/ffmpeg-ops/scripts/cut-from-edl.py edit.json --execute -o final.mp4

Rules that make this work (from the Fable launch-video pipeline): cuts must land in silence; the model reasons over transcripts, not frames; after cutting, re-transcribe the output to verify (no filler words survived, no words clipped). Full architecture, EDL schema, verification loop: references/edit-as-code.md.

Color grading

# Apply a .cube LUT (tetrahedral = highest quality interpolation):
ffmpeg -i in.mp4 -vf "lut3d=file=grade.cube:interp=tetrahedral" \
  -c:v libx264 -crf 18 -c:a copy graded.mp4

# Generate a family of grade candidates + an HTML still-chooser:
python skills/ffmpeg-ops/scripts/gen-luts.py --variants all --out-dir work/luts \
  --previews in.mp4

The human picks the grade. Generate variants, render preview stills, present a chooser — never auto-select a look. Grading is a taste call; the agent's job is the lattice math and the apply command. LUT format, log-footage normalization (S-Log3/V-Log → Rec.709), curves/eq safe ranges, checking work with ffmpeg's built-in scopes (waveform/vectorscope): references/color-grading.md. The 25-look recipe catalog — film stocks (Kodachrome, CineStill halation, Technicolor, Eterna), signature grades (Mad Max, Fincher, Matrix, BR2049, Amélie…), era/genre moods, Sin City selective color — every chain build-validated, plus the Hald-CLUT match-any-look workflow and scope-matching ladder: references/look-recipes.md. Pipeline correctness (pix_fmt, HDR→SDR tonemapping, range/matrix tagging): references/color-hdr.md.

Quality gates

# VMAF/SSIM/PSNR verdict on an encode (exit 10 = below threshold -> branch on it):
python skills/ffmpeg-ops/scripts/quality-compare.py reference.mp4 encoded.mp4 \
  --metrics ssim,psnr
python skills/ffmpeg-ops/scripts/quality-compare.py reference.mp4 encoded.mp4 \
  --metrics vmaf --min-vmaf 90 --json | jq '.data.vmaf'

VMAF ≥ 93 at 1080p ≈ visually transparent; 80-93 = noticeable on inspection. Side-by-side visual A/B (hstack), metric interpretation, encode-ladder tuning: references/quality-metrics.md.

Scripts

All eleven follow the Skill Resource Protocol: --help with examples, stdout = data only, --json envelopes (claude-mods.ffmpeg-ops.*/v1), semantic exit codes (0 ok, 2 usage, 3 input missing, 4 invalid input, 5 missing dependency, 7 ffmpeg unavailable, 10 domain finding).

| Script | Job | Worked invocation | |---|---|---| | probe-media.py | Normalized inspection, keyframe proximity, --doctor triage (hazard → fix command, exit 10) | probe-media.py --doctor in.mp4 | | capability-scan.sh | What can THIS ffmpeg build do (proof-encodes hw encoders; --quick skips) | capability-scan.sh --json \| jq '.data.encoders' — exit 10 = a listed encoder failed verification | | quality-compare.py | VMAF/SSIM/PSNR gate | quality-compare.py ref.mp4 enc.mp4 --min-vmaf 90 — exit 10 = below threshold | | loudnorm-scan.py | Two-pass loudnorm: measures pass 1, emits exact pass-2 filter | loudnorm-scan.py -I -16 in.mp4 --json \| jq -r '.data.pass2_filter' | | detect-segments.py | Silence/scene boundaries as JSON segments (STT chunking, dead-air cuts, shot splits) | detect-segments.py --scenes --json in.mp4 \| jq '.data.segments' | | cut-from-edl.py | EDL JSON → validated cuts + concat (dry-run by default) | cut-from-edl.py edit.json --execute -o final.mp4 | | make-chapters.py | Scene/silence points (or explicit JSON) → embedded chapters / YouTube text / WebVTT | make-chapters.py --from-scenes --media talk.mp4 --write chaptered.mp4 | | smart-compress.py | Fit a size cap: computed two-pass bitrate, auto audio/downscale, size-verified (exit 10 = still over) | smart-compress.py --target 25MB video.mp4 | | make-sprites.py | Scrub-preview sprite sheets + WebVTT thumbnail track (#xywh) | make-sprites.py --interval 5 video.mp4 | | gen-luts.py | Emit .cube grade variants (+ --previews still chooser) | gen-luts.py --variants warm_filmic,punchy --out-dir luts/ | | verify-commands.sh | Staleness verifier: --offline structural (CI), --live checks docs against the installed build | verify-commands.sh --live — exit 10 = doc drift, 7 = no ffmpeg |

References

Load on demand — one concept per file:

| Reference | Load when | |---|---| | encoding.md | Choosing codec/CRF/preset, two-pass, social platform targets, archival | | hardware-accel.md | NVENC/QSV/AMF/VideoToolbox/VAAPI flags, quality caveats, detection | | filtergraph.md | Any -filter_complex, labels/chains/split, speed ramps, xstack | | trim-concat.md | Cut accuracy, keyframes, concat selection, segment removal | | edit-as-code.md | Multi-cut edits, EDL schema, transcript-driven editing, verify loop | | audio.md | Loudness, mixing, channel layout, audio repair | | stt-whisper.md | Whisper/WhisperX prep, chunking, transcript JSON, summarisation pipeline | | subtitles.md | Burn vs soft, styling, extraction, format conversion | | color-grading.md | LUTs, .cube format, log normalization, scopes, grade workflow | | look-recipes.md | 25-look catalog (film stocks, signature movie grades, era/genre moods), Hald-CLUT extraction, scope-matching | | color-hdr.md | pix_fmt, HDR→SDR tonemap, BT.601/709 tagging, 10-bit | | quality-metrics.md | VMAF/SSIM interpretation, visual A/B, ladder tuning | | streaming-hls.md | HLS/DASH packaging, ABR ladders, live restream | | images-gif.md | GIF quality, sprite sheets, dataset frame extraction | | restoration.md | Deinterlace, denoise, deband, stabilize, audio cleanup | | analysis-validation.md | Corruption checks, hashing, metadata stripping, untrusted uploads | | capture-devices.md | Screen/webcam capture per OS (gdigrab/dshow, avfoundation, x11grab) | | error-decoder.md | An ffmpeg command failed with a cryptic message — symptom → cause → fix | | visualization.md | Waveform/spectrogram videos, audiograms, comparison grids | | cookbook.md | Any concrete command: the full recipe set by job (convert, cut, transform, overlay, audio, STT prep, images, diagnostics, yt-dlp, test sources) |

Assets: encoding-presets.json (recipe data incl. date-stamped social targets), hls-ladder.json (ABR ladder), edl-schema.json (the cut-from-edl.py contract).

Self-test

bash skills/ffmpeg-ops/tests/run.sh   # offline suite; synthesizes fixtures via lavfi

Structural assertions always run; media round-trips run only when ffmpeg is on PATH (loud skip otherwise — never a silent false-clean).