Claude Skills by dmmdea

Use for any FFmpeg/ffprobe work in the video pipeline — probing a media file, cutting/trimming/concatenating, retiming, scaling for YouTube 16:9 or Shorts/Reels 9:16, thumbnails, NVENC or CPU encoding choices, loudness normalization (-14/-16 LUFS), silence/dead-air removal, ducking music, burning or muxing subtitles, drawtext lower thirds, alpha overlays, GIF/WebP, image sequences, stream mapping/metadata/timecode, ffmpeg quoting in PowerShell/cmd/Git Bash, progress and exit-code handling, an...

Use when the user wants to edit, resize, crop, composite, retouch, add text/titles, remove a background, add a drop shadow or outline, batch-convert or export images with GIMP 3.2 — or any task that means driving GIMP programmatically: gimp-console batch mode, Python (GObject-introspection Gimp/Gegl), Script-Fu, the gimp-mcp server, GEGL filters, YouTube thumbnails / social-size exports, PNG/JPEG/WebP/AVIF/TIFF/PDF export. Triggers on "GIMP", "gimp-console", "python-fu", "script-fu", "make a ...

Use when setting up the "local-offload" harness on a Windows machine — a free local Gemma-4 cascade that lets a coding agent delegate short-context grunt work (summarize / classify / extract / triage) so those tokens never hit the cloud context. Cross-vendor: NVIDIA (CUDA, ≥8GB), AMD Radeon incl. RDNA3 iGPUs like the 780M/gfx1103 (Vulkan, native Windows), or CPU-only fallback. Builds a pinned llama.cpp + llama-swap stack, pulls the Gemma-4 QAT family (E2B/E4B/26B-A4B) + EmbeddingGemma, builds...

Drive a local ComfyUI render server from the shell, with a durable record of every run the server itself forgets. Trigger phrases: `submit a comfyui graph`, `why can't comfyui see my model`, `how long did that render take`, `what made this output file`, `check what this loader accepts`, `use comfyui`, `run comfyui`.

The llama-swap operations console — durable history, drain-aware control, and measurement commands with three specific guarantees: keep-set unloads are refused statically by id AND alias (never from server ttl), `--drain` fails closed when slot state is unreadable, and fit/ctx refuse to answer inside their uncertainty band instead of guessing. Trigger phrases: `what models are loaded`, `free up VRAM`, `unload a model safely`, `check llama-swap`, `is the memory stack ok`, `will this model fit`...