Separate Process From Sensitive Data When Automating

When you build an agent or skill to automate a form-heavy workflow, keep the procedure and the sensitive values in strictly separate places. The reusable automation should describe how to navigate, click, and validate — never the identity numbers, credentials, or personal details it operates on. Those live in a local, un-versioned worksheet or a secrets manager fetched at runtime, so the shareable logic stays free of anything that could leak.

Draw a hard line at actions that must stay human. Logging in, entering passwords, retrieving emailed codes, checking an “I agree” box, and clicking a native file picker are all points where the agent should stop and hand back control. Let it drive the tedious middle and pause before any irreversible or legally binding step.

Expect a real form to fight you. Framework-driven inputs often ignore a plain programmatic click and require dispatching the same event sequence a human would trigger; setting a value may need the native setter plus explicit input/change events to register. Budget for a debug cycle per subtle gotcha — cascading fields that silently rewrite each other, format rules that only surface as server errors — and encode each hard-won fix as a durable note so the next run doesn’t relearn it.