1 files changed, 36 insertions, 16 deletions
diff --git a/prompts/skills/fujifilm-video-stabilization/SKILL.md b/prompts/skills/fujifilm-video-stabilization/SKILL.md
index 330be1a..026b4e9 100644
--- a/prompts/skills/fujifilm-video-stabilization/SKILL.md
+++ b/prompts/skills/fujifilm-video-stabilization/SKILL.md
@@ -1,12 +1,12 @@
 ---
 name: fujifilm-video-stabilization
-description: "Stabilize shaky Fujifilm X100V videos (DSCF*.MOV) using ffmpeg's 2-pass vidstab filter. Handles pass 1 (motion detection → .trf), pass 2 (transform + unsharp), cleanup of temp files, and batch processing of multiple clips. Use when asked to stabilize video, reduce camera shake, smooth handheld footage, or process Fujifilm X100V MOV files. Triggers on: stabilize video, reduce camera shake, vidstab, ffmpeg stabilization, Fujifilm X100V, smooth footage, handheld video."
+description: "Stabilize shaky Fujifilm X100V videos (DSCF*.MOV) using ffmpeg's 2-pass vidstab filter and encode compact 4K HEVC MP4 output. Handles motion detection, transform and sharpening, Intel VAAPI acceleration with a software fallback, cleanup, verification, and batch processing. Use when asked to stabilize video, reduce camera shake, smooth handheld footage, compress stabilized video, or process Fujifilm X100V MOV files. Triggers on: stabilize video, reduce camera shake, vidstab, ffmpeg stabilization, Fujifilm X100V, compress video, smooth footage, handheld video."
 ---
 
 # Fujifilm X100V Video Stabilization
 
 2-pass ffmpeg vidstab workflow to remove camera shake from Fujifilm X100V `.MOV` clips.
-Outputs stabilized `.mkv` files; cleans up temp `.trf` files when done.
+Outputs compact 4K HEVC `.mp4` files; cleans up temp `.trf` files when done.
 
 ## When to Use
 
@@ -18,7 +18,8 @@ Outputs stabilized `.mkv` files; cleans up temp `.trf` files when done.
 ## Required Tools
 
 - `ffmpeg` with `libvidstab` enabled (verify: `ffmpeg -filters 2>/dev/null | grep vidstab`)
-- Input location: `/home/paul/syncthing/Documents/Inbox/Fujifilm.Videos/` (typical)
+- Intel VAAPI HEVC support through `/dev/dri/renderD128` (preferred)
+- Input location: `/home/paul/syncthing/Documents/Inbox/Fujifilm/` (typical)
 
 ## 2-Pass Workflow
 
@@ -34,18 +35,32 @@ ffmpeg -y -i DSCF7259.MOV -vf vidstabdetect=result=DSCF7259.trf -f null -
 - `-f null -` discards output; only the `.trf` sidecar is produced.
 - Run pass 1 for all clips **before** starting pass 2 (they can run in parallel if CPU allows).
 
-### Pass 2 — Stabilize & Sharpen
+### Pass 2 — Stabilize, Sharpen, and Compress
 
-Apply transforms and add a mild unsharp pass (FFmpeg's own recommendation):
+Apply transforms, add a mild unsharp pass, and encode directly to a compact HEVC MP4:
 
 ```bash
-ffmpeg -y -i DSCF7259.MOV \
-  -vf "vidstabtransform=input=DSCF7259.trf,unsharp=5:5:0.8:3:3:0.4" \
-  stabilised-DSCF7259.mkv
+ffmpeg -y -vaapi_device /dev/dri/renderD128 -i DSCF7259.MOV \
+  -map 0:v:0 -map "0:a:0?" \
+  -vf "vidstabtransform=input=DSCF7259.trf,unsharp=5:5:0.8:3:3:0.4,format=nv12,hwupload" \
+  -c:v hevc_vaapi -rc_mode CQP -global_quality 34 -tag:v hvc1 \
+  -c:a aac -b:a 128k -movflags +faststart stabilised-DSCF7259.mp4
 ```
 
 - `unsharp=5:5:0.8:3:3:0.4` — luma and chroma unsharp with mild strength (0.8 / 0.4); compensates for the slight softening vidstab introduces.
-- Output is named `stabilised-<original>.mkv` to keep it alongside the source.
+- VAAPI performs the expensive 4K HEVC encode on the Intel GPU.
+- Quality 34 reduces typical stabilized videos by roughly 80% while retaining 4K resolution. Use 30–32 when higher quality is more important than size.
+- AAC audio and `hvc1` tagging provide broad MP4 compatibility. `faststart` makes remote playback begin sooner.
+- Output is named `stabilised-<original>.mp4` to keep it alongside the source.
+
+If VAAPI is unavailable, use software HEVC encoding instead (this is substantially slower):
+
+```bash
+ffmpeg -y -i DSCF7259.MOV -map 0:v:0 -map "0:a:0?" \
+  -vf "vidstabtransform=input=DSCF7259.trf,unsharp=5:5:0.8:3:3:0.4" \
+  -c:v libx265 -preset medium -crf 28 -tag:v hvc1 \
+  -c:a aac -b:a 128k -movflags +faststart stabilised-DSCF7259.mp4
+```
 
 ### Cleanup
 
@@ -67,12 +82,15 @@ for f in *.MOV; do
 done
 wait
 
-# Pass 2 — serial (CPU-heavy; one at a time is fine)
+# Pass 2 — serial to avoid overloading the GPU
 for f in *.MOV; do
   base="${f%.MOV}"
-  ffmpeg -y -i "$f" \
-    -vf "vidstabtransform=input=${base}.trf,unsharp=5:5:0.8:3:3:0.4" \
-    "stabilised-${base}.mkv"
+  ffmpeg -y -vaapi_device /dev/dri/renderD128 -i "$f" \
+    -map 0:v:0 -map "0:a:0?" \
+    -vf "vidstabtransform=input=${base}.trf,unsharp=5:5:0.8:3:3:0.4,format=nv12,hwupload" \
+    -c:v hevc_vaapi -rc_mode CQP -global_quality 34 -tag:v hvc1 \
+    -c:a aac -b:a 128k -movflags +faststart \
+    "stabilised-${base}.mp4"
 done
 
 rm -f *.trf
@@ -83,14 +101,16 @@ rm -f *.trf
 Check duration and file size of stabilized clips:
 
 ```bash
-for f in stabilised-*.mkv; do
+for f in stabilised-*.mp4; do
   echo -n "$f: "
   ffprobe -v error -show_entries format=duration,size -of csv=p=0 "$f"
+  ffmpeg -v error -i "$f" -f null -
 done
 ```
 
 ## Notes
 
-- **Syncthing caveat**: the `Fujifilm.Videos` folder is synced. If files vanish mid-session, Syncthing may have moved them after detecting completion. Check `~/.local/share/syncthing/` or other sync peers.
+- **Syncthing caveat**: the `Fujifilm` folder is synced. If files vanish mid-session, Syncthing may have moved them after detecting completion. Check `~/.local/share/syncthing/` or other sync peers.
 - **Quality**: default vidstab params produce good results for typical X100V handheld footage. For extremely shaky video, add `shakiness=10:accuracy=15` to the pass 1 filter.
-- **Output size**: stabilized `.mkv` files are re-encoded; expect ~10–15 MB/s for X100V clips at the default encoder quality (crf 23).
+- **Output size**: HEVC quality 34 reduced a 334 MB stabilized test clip to 58 MB without changing its 3840×2160 resolution or duration.
+- **Avoid intermediate encodes**: encode the stabilized MP4 directly from the camera `.MOV`. Do not first create a VP9 MKV and then transcode it, because that wastes time and adds generation loss.