Claude's Follow-a-Plan feature needs approval gates — here's the pattern that prevents runaway execution

Claude's Follow-a-Plan feature is powerful. Maybe too powerful. It'll execute a 20-step plan without stopping to check if step 3 went sideways. That's great for demos, terrible for production.

I learned this the hard way when my agent decided to "optimize" my entire codebase structure based on a misunderstood requirement. It followed its plan perfectly — straight into a 4-hour rollback session.

The solution isn't to disable Follow-a-Plan. It's to build approval gates that catch problems before they cascade.

The Checkpoint Pattern

Instead of letting Claude execute plans end-to-end, inject approval checkpoints at critical decision points. Here's the pattern:

PLAN EXECUTION RULES:

1. Execute steps 1-3 automatically
2. STOP at any step that:
   - Modifies production systems
   - Deletes or moves files
   - Spends money
   - Changes external integrations

3. Present checkpoint summary:
   - What you've completed
   - What you want to do next
   - What could go wrong
   - How to rollback if needed

4. Wait for explicit approval: "proceed", "modify", or "stop"

This isn't about micromanaging your agent. It's about catching the moment when a good plan meets bad reality.

The Risk Ladder

Not every step needs approval. Build a risk ladder that scales oversight with potential damage:

• Green zone: Read-only operations, research, analysis — execute freely
• Yellow zone: File modifications, API calls, configuration changes — checkpoint every 3-5 steps
• Red zone: Deletions, production changes, external payments — approval required for each step

The key is making the risk assessment explicit in your agent's instructions, not hoping it'll figure out what's dangerous.

Implementation That Actually Works

Add this to your agent's system prompt:

When following a multi-step plan:

- Execute green zone steps continuously
- At yellow/red zone steps, output:
  
  ## CHECKPOINT: [Step X of Y]
  **Completed:** [brief summary]
  **Next:** [what you want to do]
  **Risk:** [what could go wrong]
  **Rollback:** [how to undo if needed]
  
  Waiting for approval to proceed...

This forces the agent to think through consequences before acting, not after.

Pro tip: Log every checkpoint decision. When something goes wrong (and it will), you'll have a clear audit trail of what the agent was thinking and what you approved.

The goal isn't to slow down your agent — it's to prevent the kind of mistakes that kill trust in autonomous execution. Better to pause for 30 seconds at a checkpoint than spend 4 hours cleaning up a runaway plan.

Your agent should be confident enough to act, but humble enough to ask. Approval gates give you both.