Image read cache

Skill ParthJadhav/image-read-cache/image-read-cache

Agent Skill that caches LLM image descriptions as XMP metadata inside image files, reducing token usage by ~92% on repeated reads. Works with 30+ compatible agents.

Install

npx -y skills add ParthJadhav/image-read-cache --skill image-read-cache

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

One thing to look at

8 stars8 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

Optimizes image reading by caching LLM-generated content as XMP metadata inside image files. Use BEFORE reading any image file (.png, .jpg, .jpeg, .webp, .gif, .bmp). Reduces token usage by ~90% on repeated image reads. Triggers on: image, screenshot, picture, photo, .png, .jpg, .jpeg, read image, view image, look at image.

SKILL.md

3.2 KB, as published. Nobody here has run it

Image Cache Skill

Purpose

Avoid sending the same image to the LLM repeatedly. Instead, cache what the LLM produced the first time it read the image, and return that cached content on subsequent reads. The cached content is stored as XMP metadata directly inside the image file.

When to use

Before EVERY image file read (.png, .jpg, .jpeg, .webp, .gif, .bmp), run the cache check first.

Workflow

Step 1: Check cache before reading

python3 <skill-dir>/scripts/check_cache.py "<absolute-image-path>"

If output starts with CACHED: — the rest of the output is the exact content the LLM produced last time it read this image. Use it directly as if the image was just read. Do NOT read the actual image file. Skip to Step 3.

If output is NO_CACHE — no cached content exists, or the image has changed since it was cached. Proceed to Step 2.

Step 2: Read image and cache the result

Read the image normally using your file-read tool (base64 to the LLM).

After the image has been processed, cache what you received. The content to cache is EXACTLY what the file-read tool returned to you — the raw tool result, unchanged. Do not summarize, reformat, restructure, or add commentary. The cached version must be a faithful replica of the original tool output so that future reads produce identical results.

Since base64 image data cannot be stored in XMP metadata, you must instead produce a comprehensive text representation of everything visible in the image. This text IS the cache — anything you omit will be invisible on future reads. Capture:

All visible text exactly as written (OCR-accurate)
Layout structure and spatial relationships (what is where)
Every UI element, color, shape, icon, and visual detail
All data: numbers, labels, chart values, table contents
Context clues: window titles, URLs, timestamps, filenames

Cache it:

python3 <skill-dir>/scripts/write_cache.py "<absolute-image-path>" "<your-complete-text-representation>"

Step 3: Continue with task

Use the content (cached or fresh) to answer the user's question.

Important rules

ALWAYS check cache before reading an image. It costs <1 second and saves thousands of tokens.
ALWAYS write back after a fresh image read. The next read becomes free.
The cache includes a file hash. If the image file changes (re-saved, re-exported, new screenshot), the cache auto-invalidates and returns NO_CACHE.
If write_cache.py fails (read-only file, permissions), continue normally. Caching is best-effort.
If the user explicitly asks to "re-examine", "look again at", or "re-read" the image, SKIP the cache check and read fresh.
Do NOT cache images the user is actively editing or generating (e.g., mid-workflow screenshots). Only cache stable assets.