The simplest reliable agent workflow is still:

1. Plan.
2. Implement.
3. Verify.

The point is not ceremony. The point is making the agent show its understanding before it changes files and making success mechanically checkable afterward.

The Pattern

Goal:
Add validation for expired invite tokens.

Plan:
First inspect the token validation path and the closest tests. Summarize the
smallest implementation path before editing.

Implement:
Write the failing regression test first. Then make the smallest production
change that passes it.

Verify:
Run bun test:unit -- src/lib/server/invitations.test.ts and bun run lint.

Stop:
If a command fails for an unrelated reason, stop and report the command, output,
and likely cause instead of changing unrelated files.

This works in Cursor Agent mode, Cursor Cloud Agents, Claude Code, and most command line agent workflows.

Why It Works

Planning catches misunderstandings while the cost is still low. Implementation keeps the agent inside the task. Verification turns “looks good” into a binary check.

The stop condition matters as much as the success condition. Without it, agents will often route around failures by editing adjacent code, weakening tests, or claiming partial success.

When to Skip the Full Pattern

For a tiny inline edit, a full plan is overkill. Use the pattern when the task:

• Touches more than one file.
• Has security, data, or user-facing risk.
• Requires tests or generated artifacts.
• Could be solved in more than one reasonable way.
• Will run in a cloud, command line, or automated context.

If you cannot name the verification command, you probably do not have a complete task yet.