diff --git a/docs/superpowers/specs/2026-07-26-jellyfin-hw-transcoding.md b/docs/superpowers/specs/2026-07-26-jellyfin-hw-transcoding.md
new file mode 100644
index 0000000..af4e585
--- /dev/null
+++ b/docs/superpowers/specs/2026-07-26-jellyfin-hw-transcoding.md
@@ -0,0 +1,108 @@
+# Intel Quick Sync hardware transcoding for Jellyfin on jupiter
+
+## Problem
+
+Jellyfin streams stutter on jupiter whenever a client needs a transcode
+(unsupported codec/container, bitrate cap, or a client that can't
+direct-play). Transcoding currently runs entirely on CPU.
+
+jupiter's Intel iGPU is already usable at the OS level:
+
+- `hardware.graphics.enable = true` with `intel-media-driver` (the `iHD`
+  VAAPI driver) is configured in
+  `machines/jupiter/hardware-configuration.nix:20-27`.
+- The commented-out `i915.force_probe = "9a49"` kernel param there
+  corresponds to a Quick-Sync-capable Intel UHD iGPU, confirming the
+  hardware supports it.
+
+But `modules/environments/jellyfin/default.nix` never grants the
+`jellyfin` systemd service access to `/dev/dri`, so Jellyfin has no path
+to the GPU and silently falls back to software transcoding.
+
+## Goal
+
+Give the Jellyfin service access to the iGPU's VAAPI render node, so
+Quick Sync can be enabled in Jellyfin's own dashboard and transcodes are
+offloaded from the CPU.
+
+## Non-goals
+
+- Remote/external access or reverse-proxy tuning.
+- General CPU/RAM headroom review of jupiter.
+- A toggle option (`my.profiles.jellyfin.hardwareAcceleration.enable`) —
+  jupiter only has the one iGPU, so this is hardcoded on rather than
+  made configurable.
+
+## Design
+
+### NixOS change (declarative)
+
+In `modules/environments/jellyfin/default.nix`, inside the existing
+`config = lib.mkIf cfg.enable { ... }` block, grant the systemd service
+supplementary access to the `video` and `render` groups (the groups that
+own `/dev/dri/card*` and `/dev/dri/renderD*`):
+
+```nix
+systemd.services.jellyfin.serviceConfig.SupplementaryGroups = [
+  "video"
+  "render"
+];
+```
+
+This is additive to the existing `systemd.services.jellyfin.after = [
+"network-online.target" ];` block already in the file — both apply to
+the same service.
+
+Also add `libva-utils` to `environment.systemPackages` (or scoped to
+this module) so `vainfo` is available on jupiter to verify the driver
+loads correctly.
+
+### Manual step (not declarative)
+
+Jellyfin stores its transcoding/hardware-acceleration choice in its own
+internal `encoding.xml`, which the NixOS module does not expose as an
+option. After deploying the Nix change, one-time manual configuration in
+the Jellyfin dashboard is required:
+
+1. **Dashboard → Playback**.
+2. Hardware acceleration: **Intel QuickSync (QSV)**.
+3. VA-API device: `/dev/dri/renderD128`.
+4. Enable hardware decoding for the codecs your library actually uses
+   (H264 at minimum; HEVC/VP9 depending on iGPU generation).
+5. If any HDR content exists in the library, enable tone-mapping — this
+   is one of the more CPU-expensive operations Quick Sync can offload.
+
+## Verification
+
+Build-time (from the Mac, no SSH needed):
+
+```
+nix eval '.#nixosConfigurations.jupiter.config.systemd.services.jellyfin.serviceConfig.SupplementaryGroups' \
+  --extra-experimental-features 'nix-command flakes'
+```
+
+Expect `[ "video" "render" ]`.
+
+On jupiter after `sudo nixos-rebuild switch --flake '.#jupiter'`:
+
+```
+systemctl status jellyfin
+journalctl -u jellyfin -n 50 --no-pager
+vainfo
+```
+
+`vainfo` should list the `iHD` driver and print supported VAEntrypoints
+(VLD decode / encode profiles for H264/HEVC).
+
+Functional check: play a file on a client that forces transcoding (or
+force it manually via Jellyfin's playback quality setting), then in
+Jellyfin's dashboard **Activity/Now Playing** panel confirm the
+transcode reason and check that CPU usage on jupiter (`htop`) stays low
+during playback rather than pegging a core — Quick Sync offload should
+show up as low CPU, some GPU (`intel_gpu_top`) activity instead.
+
+## Open items
+
+- Exact supported codec list depends on the iGPU generation (device ID
+  `9a49`) — confirm via `vainfo` output once run, and enable only the
+  hardware decode paths it actually reports.