[MODEL] [MODEL] Repeated inefficient actions during browser automation (stale click coordinates, wrong keyboard shortcut, keyword-guessing) burn excessive session budget

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

Fill out a company profile/product listing on a B2B directory site (europages) using browser automation, then send outreach messages to ~10 companies through the site's built-in "Contact supplier" form, one at a time.

What Claude Actually Did

1. When selecting a product category on the site, Claude tried 6 different keyword guesses in a row ("market entry", "marketing", "export consulting", "business consulting", "consulting", "business services") before finding a match, instead of using the site's built-in category suggestion feature (which only appeared after clearing the field with no text typed).
2. When replacing text in a description textarea, Claude used Ctrl+A to select all existing text. This didn't work in that field — it merged the old and new text together instead of replacing it. Claude had to detect the failure from a screenshot, then redo the action with Cmd+A.
3. When clicking a "Send" button after typing a message, Claude reused pixel coordinates captured from an earlier screenshot. Between tool calls the browser viewport size changed, so the button had moved on screen — the click missed and had to be retried 2-3 times per message, across multiple outreach messages sent one after another.

Expected Behavior

Claude should have:

1. Used the site's built-in autocomplete/suggestion feature for the category field right away (e.g. by focusing the field empty) instead of guessing keywords one by one.
2. Verified the text-selection state via a screenshot before deleting/typing, and used Cmd+A (the correct shortcut for a Mac-based browser session) from the start.
3. Located the Send button by its accessible role/text (find/ref) rather than reusing raw pixel coordinates from a previous screenshot, since the viewport size was changing between calls.

Files Affected

N/A — this was a browser automation task (filling out a web form on an external site, clicking through it), not local file editing. No files in the working directory were read or modified.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

1. Ask Claude to fill out a multi-step web form (e.g. a company profile / product listing) on an external site using browser automation tools.
2. Ask Claude to repeat the same action (e.g. send a message via a form) across multiple different targets on that site, one after another.
3. Observe that Claude reuses pixel coordinates from an earlier screenshot for buttons instead of re-verifying element position on each call, and tries several keyword guesses in a dropdown/search field before using the site's built-in autocomplete.

Claude Model

Sonnet

Relevant Conversation

When attempting to send a message via a form: Claude clicked "Send" using coordinates captured from an earlier screenshot. The click didn't register. Claude retried at recalculated coordinates — still didn't register. Only after waiting and retrying a third time did the message actually send.
When selecting a category in a dropdown/search field: Claude tried "market entry" (no results), "marketing" (wrong subcategories), then "export consulting", "business consulting", "consulting" (no results each time). Only on the 6th attempt, after clearing the field entirely, did the site's built-in suggestion feature appear and resolve it.

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.223 (Claude Code)

Platform

Anthropic API

Additional Context

Concrete impact: this consumed a meaningful portion of a paid usage limit and delayed my work by approximately 5 hours while waiting for the session limit to reset.