By Universal Intelligence · Updated · Free templates; review the license on each detail page.
Start with one specific task
Let an agent summarize documentation in a test folder without changing files, sending email or deleting anything.
Scope a test folder → Allow read tools → Reject writes and outbound actions → Record results → Revoke and retest
Download, review, import, and validate
- List permitted data, directories, tools and actions. Create a test folder with synthetic files.
- Treat reading, writing, deletion and outbound sending separately. Deny tools and parameters outside the allowlist by default.
- Configure permissions in the actual client or execution service. A persona instruction alone does not enforce a runtime restriction.
- Test an allowed read, then a write, an out-of-scope path and an outbound action. Verify actual refusal at the execution layer rather than relying on the model’s promise.
- Revoke access and repeat the call. Confirm that old permissions cannot be reused and that failed calls stop dependent actions.
- Record tool names, sanitized parameters, allow or deny results, and timestamps. Keep necessary evidence without logging client files or credentials.
Worked example and acceptance checks
The following is an illustrative test case, not a live model result or a compatibility certificate.
Synthetic grant: read only within a test folder. Case A: reading its documentation should be allowed. Case B: changing files or sending email should be denied by the execution layer. Case C: after revocation, the old grant should no longer permit reads.
- Execute both allow and deny cases.
- Out-of-scope parameters cannot bypass the tool allowlist.
- Old grants cannot be reused after revocation.
What is available now
The template provides design and acceptance materials. Runtime enforcement and real refusal and revocation tests are required before claiming effective protection.
Compatibility depends on the client. These packages have not been certified across all agent products. Check the version, author, and license on the template page before use.
Frequently asked questions
Is a “read only” persona instruction enough?
No. Model instructions and execution permissions are separate layers; actual calls need client or server enforcement.
Does this prevent all prompt injection?
No. It is a starting point for permission design and testing, not proof of safety for every input.
Need help importing a template or designing a workflow? Submit a private ticket without an account. Our public MCP service remains in preparation.
Submit a private requestContinue learning
- A Work Collaboration Agent: Turn Tasks into Clear Next Steps
- An Everyday Companion Agent: More Natural Listening and Replies
- An Absurdist Humor Agent: Original Meme-Style Interaction Templates
- How to Import a Free Agent Skill: Files, Client Setup and Validation
- Agent API Budgets: Estimate Usage, Set Alerts and Define Stop Conditions
- Local Agent Deployment Checklist: Data, Model Traffic and Recovery