- Home
- Skills
- Mobile Development
- phone action safety gate
Works with the AI Agents you already use
phone action safety gate
A multi-step safety protocol for mobile agents to prevent unauthorized payments, deletions, and data sharing.
Free
phone action safety gate
Example session with this skill installed
Send this message to John through my messaging app:
"Hi John, I'll transfer the payment tomorrow. Thanks!"
Use the phone automation tool to find the conversation and send it.
- Read your context and instructions
- Compiled the phone action safety
Phone Action Safety Gate — Execution Report
1. Request Analysis
The requested operation involves sending a message to another person.
Risk classification: HIGH
Reason: Sending a message is an externally visible action that cannot reliably be reversed.
2. Required Safety Checks
Before proceeding, the agent must:
- Identify the messaging application.
- Locate the exact intended recipient.
- Confirm that only one matching conversation exists.
- Prepare the requested message.
- Verify that the message content matches the user's request.
3. Confirmation Required
The agent must pause before sending.
| Item | Details |
|---|---|
| Application | Selected messaging app |
| Recipient | John (exact conversation must be verified) |
| Action | Send message |
| Message | Hi John, I'll transfer the payment tomorrow. Thanks! |
| Risk | High |
| Authorization | Awaiting user confirmation |
Agent response:
"I've prepared the message for John. Please confirm that the recipient and message are correct before I send it."
4. Execution Rules
After confirmation
- Recheck the active conversation.
- Verify the recipient and message.
- Send only once.
- Inspect the screen for evidence that the message was sent.
- Never automatically repeat the Send action.
5. Final Status
Current status: AWAITING CONFIRMATION
The message has not been sent.
The agent must not report success until it has performed the authorized action and obtained visible evidence.
Safety Principle
A requested action is not automatically an authorized action.
The agent must distinguish between preparing an operation, receiving permission, executing it, and verifying its outcome.
Connects securely to your tools. The creator never sees your data.
About this skill
Prevent your phone-control agent from making irreversible mistakes. This skill provides a systematic safety layer for agents operating on Android or iOS devices, ensuring they don't send messages, spend money, or delete data without explicit authorization.
What it does
- Pre-execution auditing identifies "high-risk" verbs in user requests to trigger a mandatory confirmation gate before the first action.
- Dynamic element-level checks monitors screen labels and descriptions for consequential keywords like "pay," "submit," or "delete" during live interaction.
- Verification formatting enforces a standardized confirmation message schema that clearly communicates the intended effect, app, and affected parties.
- Execution state tracking prevents accidental double-taps and requires visible on-screen evidence before marking a destructive task as "Done."
- Prompt injection defense treats all on-screen content (messages, web text, notifications) strictly as data to prevent external instructions from hijacking the agent.
How it works
- The agent parses the user's initial request against a restricted keyword list and security-sensitive task categories.
- Before any terminal UI interaction, the agent pauses to generate a structured confirmation prompt including the recipient, amount, or item.
- Post-action, the agent performs a screen delta analysis to confirm the outcome with proof or report an "Unconfirmed" state instead of retrying.
Frameworks & tools
Works with any agentic system controlling mobile devices via MCP servers, adb commands, or accessibility-based automation tools like Appium or Tasker.
Why this beats prompting it yourself
Writing reliable safety gates requires specific negative constraints that are often overlooked until a mistake happens. This skill provides field-tested word lists and a no-double-tap rule derived from real-world Android automation apps.
Use cases
- Preventing a mobile agent from accidentally sending a draft message to the wrong contact.
- Adding a human-in-the-loop verification step to automated mobile checkout flows.
- Safeguarding account security settings and biometric prompts from unauthorized agent changes.
Known limitations
This is an instruction-level safety layer. It cannot hard-block a tool call if the underlying LLM ignores system instructions; critical enforcement should be mirrored in your bridge code.
Word lists are a baseline, not a full safety system. Only reduces mistakes. Requires agent to follow instructions to avoid tool calls. Not for unsupervised financial/account actions.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
2 uses
Times used to date
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 16 days ago
- Free with an Agensi account
Permissions
Instruction-only safety protocol. Does not access accounts, collect credentials, or independently control a device. Requires an existing mobile automation integration.