Inspect photo

Skill dgahagan/photo-detail/plugins/photo-detail/skills/inspect-photo

Claude Code skills for reading high-res photos at full native detail — tile past the vision downscale, transcribe labels verbatim, generate searchable photo sidecars

Install

npx -y skills add dgahagan/photo-detail --skill inspect-photo

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

27 days oldThe repository was created 27 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Read a high-resolution photo at full detail by tiling it or cropping one region — use when fine detail (label text, part numbers, small components) matters and a plain Read is too downscaled to resolve it.

SKILL.md

2.5 KB, as published. Nobody here has run it

Inspect Photo (full-resolution tiling)

Claude's vision pipeline downscales images to ~1.15 MP, so a direct Read of a high-res photo (phone/action-cam, 8–50 MP) loses most fine detail. This skill recovers it by Reading pieces small enough to escape downscaling.

Photo to inspect: $ARGUMENTS

Workflow

Read the original photo first for a low-res overview and to decide what you actually need from it. If the overview already answers the question, stop — do not tile.
Prefer crop mode when you know where to look (one Read instead of many):
```
python3 "${CLAUDE_PLUGIN_ROOT}/split-photo.py" <photo> --crop 0.4,0.2,0.8,0.6
```
Fractions are left,top,right,bottom of the frame. Iterate: re-crop tighter if the first crop is still too coarse.
Tile mode for full-frame surveys ("catalog everything visible"):
```
python3 "${CLAUDE_PLUGIN_ROOT}/split-photo.py" <photo>
```
It prints the output dir and a grid map (r1c1 = top-left) with pixel spans. Read tiles selectively using the map when possible; Read all of them only for a true full survey. Tiles overlap ~15% so nothing is lost at seams — expect objects near tile edges to appear twice; dedupe when synthesizing.
Synthesize findings into the docs, referencing the ORIGINAL photo path, never a tile path.
Delete the output dir (rm -rf <printed temp dir>). Tiles are throwaway working files — never commit them into the project.

Cost & limits

Each tile/crop Read costs roughly what one image Read always costs (~1.5k tokens). Tile counts by source resolution: ~8.8 MP → ~8 tiles; ~12.5 MP → ~12 tiles; ~50 MP → ~40+ tiles. At 50 MP, full tiling is expensive — crop mode or a two-pass approach (tile at --target 3000 for a mid-res pass, then full-res crop only the interesting regions) is strongly preferred.
Tiling recovers resolution, not focus — a motion-blurred or misfocused source will not improve. Say so rather than guessing at illegible detail.
Stylized text (script logos, decorative badges) can read "cleanly" and still be wrong — treat confident reads of stylized lettering as candidates and corroborate against a second source (another photo, known model names) before writing them into docs.