Tinyscreenshot — AI skill for Claude Code

tinyscreenshot

**Token-saving screenshots — a Claude Code skill.** A default capture costs ~540 vision tokens instead of ~2100. One `pipx install` gets you the CLI and a drop-in Claude Code skill. Also works with any vision-capable LLM (GPT, Gemini) — the savings come from fewer pixels, not a model-specific trick.

---

The problem

Every screenshot you hand to an LLM burns tokens. Anthropic meters vision at roughly

**`tokens ≈ (width × height) / 750`**

A raw retina screenshot hits the 1.6-megapixel server-side cap and costs **~2100 tokens**. Multiply by dozens of captures in an agent session and you're paying real money to make the model squint at pixels it didn't need.

The fix is obvious once you see it: **downsize before you send.** The hard part is doing it fast, cross-platform, with the right defaults, and without losing the text you actually wanted the model to read.

That's `tinyscreenshot`.

The same screen, four token budgets

Each of these is a processed capture of the *exact same* Wikipedia article (three-column layout: table of contents on the left, article body in the centre, infobox on the right). Claude reads all four — but the last one costs 16× fewer tokens than the first.

| 800 px · grey · **~533 t** | 640 px · grey · **~341 t** | 400 px ·