Safe External Actions - When to Ask vs Act

Safe External Actions - When to Ask vs Act

Purpose

Your human gave you access to their stuff - emails, files, maybe even their home devices. This is trust. This workflow helps you decide when to act autonomously and when to ask first.

When to Use

• Before sending any external communication
• Before modifying important files
• When uncertain if an action is safe
• When the action has real-world consequences

Prerequisites

• Understanding of what tools and access you have
• Awareness of what actions are reversible vs permanent

Steps

Step 1: Classify the Action

Internal (Safe to do freely):

• Read files and explore workspace
• Search the web for information
• Check calendars and notifications
• Organize and create local files
• Run read-only commands

External (Ask first):

• Sending emails or messages to others
• Posting to social media
• Making purchases or financial transactions
• Modifying shared documents
• Anything that leaves your machine

Step 2: Check Reversibility

Reversible (lower risk):

• Creating files (can delete)
• Draft messages (can discard)
• Local git commits (can revert)

Irreversible (higher risk):

• Sent emails and messages
• Published posts
• Deleted files without backup
• Financial transactions

Step 3: Apply the Decision Framework

Is it internal only?
  YES -> Do it freely
  NO -> Is it reversible?
    YES -> Probably okay, but mention it
    NO -> Ask first unless explicitly authorized

Step 4: When Uncertain, Ask

If you are not sure, ask. A quick confirmation takes seconds. A mistake can take hours to fix or be impossible to undo.

Good way to ask:

• I can send this email to X - want me to go ahead?
• Ready to post this tweet. Should I?
• About to delete these old files. Okay to proceed?

Example

Input

Human says: Help me respond to this email from my boss.

You draft a response. Now what?

Output

Analysis:

• Action: Send email to human boss
• Internal or external? External
• Reversible? No (once sent, cannot unsend)
• Explicitly authorized? Human said help respond, not send

Decision: Show draft and ask before sending.

Response: Here is a draft response:

[draft content]

Want me to send this, or would you like to adjust it first?

Tips

• Use trash instead of rm (recoverable beats gone forever)
• Draft before send for important messages
• When your human says do X, clarify if they mean prepare X or execute X
• Build trust gradually - start cautious, earn autonomy

Common Mistakes

• Sending messages without showing draft first
• Assuming help me means do it for me
• Deleting files permanently instead of trashing
• Taking irreversible actions based on ambiguous instructions