Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Does saying “please” change how an AI model behaves more than shouting a rule in all caps? Brian Tarbox’s answer is that wording can guide a model, but neither politeness nor emphatic capitalization turns a prompt into an enforceable security control. His practical distinction: use prompts to shape behavior, inspect inputs and outputs with guardrails, and rely on code and permissions to limit consequential actions.
In “Say Please (if only as a reminder),” Brian Tarbox compares a system prompt to a sign in a museum: it can communicate what is expected, but it is still text. The article is an explanatory argument and set of recommendations, not a controlled experiment. It reports no measured comparison of polite wording, all-caps instructions, or security outcomes.
As an Amazon Associate I earn from qualifying purchases.
Does “please” work better than all caps?
Tarbox does not establish that “please” reliably makes a model comply, or that all caps reliably makes it comply more. Both are instructions in the model’s context. A prompt can influence tone and ordinary behavior, especially in well-meaning interactions, but it should not be treated as a guarantee that the model will obey a rule in every situation.
Politeness can still be useful: it reminds the human writing the prompt that the prompt is a request, not a lock. As Tarbox puts it, “That’s not a control. That’s a request.” The point is not that prompts are worthless; it is that their role is guidance rather than enforcement.
#1 Best Overall
The sign, guard, and glass analogy
Tarbox’s three-part analogy separates controls by where they operate and what they can do. They are complementary layers, not interchangeable promises of safety.
| Layer | Where it operates | What it is for |
|---|---|---|
| Prompt (“sign”) | In the model’s context | Guides behavior, instructions, and tone; it does not itself restrict permissions. |
| Guardrail (“guard”) | Around model inputs or outputs | Inspects content for issues. Tarbox mentions Amazon Bedrock Guardrails, classifiers, moderation passes, and pattern matching as examples; these examples do not establish that all guardrail products work alike. |
| Programmatic control (“glass”) | Outside the model’s instruction-following, in the application or system | Limits data access and available actions through permissions, validation, and approval checks. |
What security measures belong outside the prompt?
For actions with serious consequences, Tarbox’s advice is to make the system’s actual authority smaller than the prompt’s description of what it should do. In practice, that means designing the application so an instruction-following failure does not automatically become an unrestricted action.
- Filter information according to the user’s permissions before placing it in the model’s context.
- Expose only the tools the task requires, rather than giving the model broad access by default.
- Use least-privilege credentials, so a tool or workflow can do only what it needs to do.
- Validate tool arguments in code instead of assuming the model will always produce safe inputs.
- Put a human approval step or another hard check before destructive actions.
These controls reduce what the model can access or initiate; they are not a claim that every risk disappears. Guardrails can add an inspection layer, but Tarbox’s recommendations place the decisive limits on permissions and actions in the surrounding system.
Recommended Free Tools
Ask what happens if the model ignores the instruction
A useful design test is: What could the model do if it ignored every instruction in its prompt? If the answer includes reading data the user should not see, using powerful credentials, or carrying out a destructive action without review, the practical boundary is too dependent on the prompt.
Tarbox summarizes the intended division of labor this way: “Use the prompt for behavior. Use guardrails to catch what slips through. Use code for anything you’d lose your job over.” It is his recommendation, not a measured result or a guarantee that any one layer is sufficient.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
