As your AI workflows become more capable — running on schedules, querying live data, chaining steps — a new question emerges: when should you review before Claude acts, and when should you let it run fully on its own?

This module gives you a practical framework for making that decision.

The Supervision Spectrum

There is a spectrum between “Claude does nothing without my approval” and “Claude does everything autonomously.” The right position on that spectrum depends on the stakes of the specific task.

Supervised	Semi-autonomous	Fully autonomous
Claude presents a plan or draft; you approve before anything happens	Claude takes action, but notifies you and you can override	Claude takes action and reports results; no approval needed
Best for: consequential actions (sending emails, making changes to data)	Best for: actions that are easy to reverse but benefit from you being aware	Best for: analysis, reporting, read-only queries, low-stakes outputs

How to Decide: The Three Questions

For any automated task, ask these three questions:

1. Is this reversible?

If Claude saves an analysis file to /analysis, you can delete it if it is wrong. If Claude sends an email to your supplier, you cannot unsend it. Irreversible actions should always have a human review step until you are very confident in the workflow’s reliability.

2. What is the blast radius if something goes wrong?

A wrong PPC analysis saved to a file: low stakes, you just discard it. A wrong bid applied directly to your live PPC campaigns: high stakes, real money. Match the level of oversight to the potential damage.

3. How many times has this workflow run correctly?

A new workflow should be supervised for the first 3-5 runs. After that, if the output is consistently reliable, you can move it to semi-autonomous or fully autonomous depending on its stakes.

A Practical Framework for Amazon Workflows

Fully autonomous — run these with no review

• Weekly PPC digest (analysis and reporting only — no changes made)
• Inventory status reports
• Order summaries and anomaly flags
• Monthly profit summaries
• Listing quality audits (read-only scan)

Semi-autonomous — run these and notify, but no action without review

• Reorder recommendations (Claude flags the need, you place the order)
• Bid adjustment recommendations (Claude calculates, you apply)
• Draft supplier emails (Claude writes, you send)

Supervised — always review before anything happens

• Any action that sends an external communication
• Any action that modifies a live system (PPC campaigns, listings)
• Any new workflow in its first few runs

Building Trust Incrementally

The right approach is to start supervised and gradually extend autonomy as the workflow earns it.

A practical progression for a new workflow:

1. Runs 1-3: Supervised. You review the output and give feedback before anything is saved or sent.
2. Runs 4-6: Semi-autonomous. Claude runs and delivers output. You review after the fact and note any corrections.
3. Runs 7+: If consistently reliable, move to fully autonomous for read-only workflows, or keep semi-autonomous for anything with external impact.

Do not skip the supervised phase. The first few runs of any workflow reveal edge cases and misunderstandings that would otherwise compound in autonomous operation.

How to Add a Review Step to a Scheduled Workflow

If you want a workflow to pause for your approval before completing, add a “draft only” instruction to the relevant step:

"[...analysis instructions...]
Save the analysis to analysis/ppc-flags-[date].md
and send me a Slack notification that the analysis is ready for review.
Do NOT take any further action until I confirm."
        

This pattern — analyze, save, notify, wait — gives you the efficiency of automation while maintaining control over consequential next steps.