Claude is screenshotting your browser to find every button. Make it stop.

Why Claude keeps reaching for the camera

Claude in Chrome clicks through a tool called computer, and that tool's own description tells the model to take a screenshot first and work out pixel coordinates from the picture. Left alone, Claude does what the tool says: look, guess a spot, click, look again.

That loop costs you twice. Every screenshot is a full round-trip — the page is captured, shipped to the model, and read like a photograph — so sessions slow to a crawl. And coordinates are a guess at a pixel position, not a hold on the thing itself: the moment the page reflows, a banner loads, or the window sits at a different width, the guess lands on the wrong element. That's the misclick you keep watching it make.

The browser already says where everything is

None of that looking is needed, because the page carries a machine-readable map of itself — the accessibility tree, the same structure a screen reader uses. Claude's browser tools can read it directly:

• find — takes a plain-language description ("the blue Submit button") and returns a reference to the element itself.
• read_page — lists every element on the page; filtered to interactive ones, it's a map of everything clickable, each with a reference.
• Refs instead of coordinates — passing that reference to the click tool hits the element wherever it currently sits. Reflow doesn't matter. The click can't drift.
• form_input and key presses — fields filled by reference, Enter and Tab and Escape sent as keys. No clicking into a box by position.
• URL deep-links — the fastest click is no click. Going straight to the settings page beats clicking through three menus, and it can't miss.

Exact, fast, and no camera. The only thing missing is a standing instruction to use them — because the tool's own default says otherwise, and the default wins until you overrule it in writing.

The fix: one standing rule

Claude Code reads a plain-text file of standing instructions at the start of every session. The global one lives at ~/.claude/CLAUDE.md and applies to every project on your machine. A rule written there doesn't need repeating, doesn't fade when the session ends, and outranks what a tool says about itself.

Open that file — create it if it doesn't exist — and paste this section in, exactly as written:

## Browser Automation — Native Navigation Only (HARD RULE)
When driving Claude in Chrome (the mcp__claude-in-chrome tools), use NATIVE navigation. Never drive by screenshots:
- Navigate by URL deep-links whenever possible instead of clicking through menus.
- Locate elements with the find tool (natural-language search) or read_page (accessibility tree — use filter: "interactive"). Both return element refs.
- Click by passing that ref to the computer tool instead of coordinates. Never take a screenshot to figure out where to click.
- Fill fields with form_input using the ref. Use key actions for Enter, Tab, Escape. Use scroll_to with a ref to bring elements into view.
- Verify each step by running find or read_page on the new state and checking the text — not by screenshotting.
- Screenshots are allowed ONLY when the deliverable itself is visual (design review, capturing proof for me) — never as the driving mechanism.
Why this rule exists: the computer tool's own description tells you to consult a screenshot to determine coordinates before clicking. Ignore that default. Screenshot-coordinate clicking is slow (a full round-trip per look) and misclicks constantly. Accessibility refs are exact and fast.

Note what the rule keeps: screenshots stay allowed when the picture is the deliverable — a design review, proof of a finished step. They're banned only as the way of finding things. That honest carve-out matters; a rule that bans a tool outright gets broken the first time the tool is genuinely needed.

Check it took

Read the file back — not what you meant to save, what's actually there:

grep -A 10 "Browser Automation" ~/.claude/CLAUDE.md

If the section prints, it's live. Every new session picks it up on its own — nothing to restart, nothing to re-prompt. Then watch the next browser task: elements found by description, clicks landing by reference, each step verified by reading the page's text. A screenshot should only appear when you asked for a picture.

Why the rule reads the way it does

Three choices make this rule hold where a passing "please stop screenshotting" doesn't:

• It names the tools. Not "browse efficiently" — the exact tool for each job: find and read_page to locate, refs to the click tool, form_input for fields. Vague rules get vague obedience.
• It names the default it overrules. The rule says out loud that the click tool's own description points the wrong way, and to ignore it. Without that line, the model drifts back to what the tool tells it every time.
• It carries its why. Slow, misclicks, refs are exact. A rule with its reason attached gets followed in cases the rule-writer never pictured; a bare command gets lawyered.

If that shape sounds familiar, it should — it's a standard operating procedure, written for a machine. The same move as stating your crawl policy instead of implying it: the written version survives the next session, the next machine, the next tool update. The unwritten version lasts until the default reasserts itself.