Lesson Content

Calibrating autonomy to task stakes.

Not all Cowork tasks warrant the same level of oversight. The right autonomy level matches the reversibility and consequence of the task:

Low-stakes, reversible tasks: Higher autonomy is appropriate. Claude can work through multiple steps with minimal check-ins. Example: synthesizing documents, drafting content, summarizing research. If something goes wrong, it is caught in the final review and easily corrected.

Medium-stakes tasks: Checkpoint-based oversight is appropriate. Define specific points in the task where Claude should confirm before proceeding. Example: large document productions, multi-step analysis. Check in at logical phase boundaries.

High-stakes, hard-to-reverse tasks: Step-by-step review is appropriate. Claude should confirm before every significant action. Example: any task involving write access to connected apps, production communications, financial calculations.

Early signals a task is drifting.

In multi-step work, errors compound – a wrong early step can invalidate later work. Watch for:

• Output that addresses a different question than the one you asked
• Scope creep: Claude is doing more than assigned, touching things you did not ask it to touch
• Format that does not match your instructions
• Confident claims about factual matters that seem inconsistent with your knowledge
• Tools or resources being used that you did not specify

Using steering input to correct course.

When you notice drift, specific steering is more effective than a generic "that's wrong":

• "You are summarizing the wrong document – I need the 2026 version, not 2025."
• "Your scope has expanded into section 4 of the report – I only need sections 1 and 2."
• "The format here is wrong – I need a table, not a narrative."

Specific steering preserves the work done so far and redirects only the errant element.

Heightened oversight for consequential tasks.

For any Cowork task that connects to real apps, produces output for external distribution, or involves data that affects real decisions: use Claude in a "show me before you do" mode. In the task assignment, include: "Before executing any action in [connected app], describe what you are about to do and wait for my confirmation." This adds a review step before each consequential action.