McKay Wrigley@mckaywrigley
Builder; Takeoff AI Academy
Browser agents for repetitive web tasksLegacy-UI automation
Representative of McKay's threads on deploying Claude with computer-use or a browser agent to handle repeat tasks — expense uploads, HR forms, PTO requests — that live behind legacy UIs.
“The agent isn't beating the API — there is no API. That's the whole point.”
How to replicate
- 1.Pick a task that has no API and a stable UI (expense portal, benefits site, old CRM).
- 2.Record a successful human run; annotate it into 5–10 discrete steps.
Prompt template
You are a browser agent completing <task> in <application>. Follow these steps exactly: <numbered list>. Before each destructive action (submit, delete, send), output PAUSE and wait for the user to type GO. If a page looks different from expected, stop and describe what you see. Do not improvise.
Pitfall
Removing the human gate too early. Legacy UIs change silently; the agent will happily submit a wrong form 50 times before you notice.
What you'll learn
- •Where browser agents beat API-based automation
- •Why PAUSE/GO checkpoints matter on destructive actions
- •How to anchor an agent with screenshots of the start state
- •When to give up and write a Zapier integration instead
#browser-agent#automation#web