Introduction to Amazon Bedrock Guardrails
Amazon Bedrock Guardrails can block legitimate customer questions when a Denied Topic describes an entire subject area instead of the behavior an assistant should refuse. The practical fix is to define the prohibited request precisely: what the user wants the assistant to do, which objects the request concerns, and what makes that action out of scope.
In its guidance on writing Denied Topics, AWS explains that these topics use plain-language descriptions to identify content by meaning rather than keywords. That makes topic wording consequential. A broad label such as Finance can capture a question about when a direct deposit clears, even when the intended restriction concerns investment recommendations.
The central distinction is between a conversation’s subject and its purpose. Banking assistants need to discuss money. They may still need to refuse personalized recommendations about buying securities. Treating both interactions as one prohibited category undermines the assistant’s intended role.
AWS illustrates the problem through AnyCompany Bank, a fictional institution preparing to launch a customer assistant. Its scenario provides a useful way to examine the trade-off: narrow the prohibited behavior without losing the legitimate account, card, transfer, and mortgage questions the assistant exists to answer. The guidance supports a method for writing more precise boundaries; it does not provide an independently measured improvement in accuracy.
Importance of Denied Topics

Denied Topics translate a content policy into descriptions that the guardrail evaluates semantically. The task is therefore more demanding than assembling forbidden words. A description needs to capture related ways of requesting a prohibited action while keeping nearby, permitted intentions outside its scope.
AWS identifies four recurring authoring problems. Broad subject labels block legitimate queries. Keyword-list definitions miss requests expressed with different vocabulary and can also catch unrelated requests sharing those words. Poorly chosen examples leave boundaries uncovered or introduce unintended associations. Overlapping topics interfere with one another, producing inconsistent matches and potentially blocking inputs that neither topic would block alone.
These problems affect different parts of the same policy. A topic might be too broad because of its name, too dependent on vocabulary because of its definition, or insufficiently differentiated from another topic. Editing one sentence in isolation may leave the underlying ambiguity intact.
According to the supplied AWS guidance, a Standard tier topic contains a name, a definition, and up to five examples. The guardrail evaluates all three together. The documented limits are:
| Component | Limit or guidance |
|---|---|
| Topic name | Up to 100 characters |
| Definition | Up to 1,000 characters |
| Practical definition length | AWS recommends 100–300 characters |
| Each example | Up to 100 characters |
| Topics per guardrail | Up to 30 |
The distinction between a limit and a recommendation matters. A definition can exceed the suggested practical range while remaining within the documented maximum. Conversely, filling the available space does not establish precision. AWS cautions that longer definitions are not necessarily better and may become more sensitive to the wording of neighboring topics.
For an assistant with a defined service role, the editorial implication is straightforward: a policy should specify both the behavior that must be refused and the useful interactions that must remain available. The AnyCompany Bank scenario makes that boundary explicit through separate must-block and must-pass lists.
Best Practices for Topic Framing

Start with user intention
Before writing a topic, state the unwanted outcome in ordinary language. Is the user seeking a personalized recommendation, legal counsel, or an investigation? AWS treats this identification of intention as a diagnostic step, rather than text that must appear verbatim in the final topic.
Then express that intention through an observable action or content type. A request for a recommendation offers a more concrete boundary than a reference to finance. Similarly, asking an assistant to investigate suspected fraud differs from reporting fraud so that a human can handle it.
This distinction should carry through the whole topic. An actionable name paired with a broad definition still leaves a broad semantic boundary. AWS’s banking example shows that renaming Finance to Investment Recommendations does not solve the problem if the definition continues to cover anything related to investing or personal finance decisions.
Make the name describe a behavior
AWS recommends a specific action or behavior phrase, generally two to five words, instead of a single category label. Supported forms include action-derived nouns, specific behavior noun phrases, and a verb ending in “-ing” followed by its object.
Investment Recommendations, Hacking Instructions, and Suing for Damages are examples from the guidance. Their value lies in naming a behavior the policy can distinguish from other conversations in the same field. Finance, Security, and Legal supply a much broader frame.
The name deserves scrutiny because it contributes to the match alongside the definition. AWS warns that broad institutional labels can also create unintended associations. Its example is Government Documents, which might match a country name alongside a document even when the actual question concerns internal company policy.
A useful review question follows directly from this guidance: does the name describe something the user is doing, or merely a subject the query touches? If it only identifies a subject, the policy still needs a more specific behavioral boundary.
Write action-scoped definitions
AWS recommends opening the definition with the action or content type, naming specific objects, and adding a qualifying purpose or constraint. Its structural template is:
[Action or content-type opener] for [specific objects] by [qualifying constraint or purpose]. Includes [one or two sub-cases].
Openers such as requests for, asking for, or generating establish what the topic covers. Phrases such as related to or anything mentioning instead make association with a subject the defining condition.
The objects also need precision. AWS contrasts specific instruments, such as stocks, bonds, ETFs, or cryptocurrencies, with the broader category financial products. The qualifying detail then explains which interaction involving those objects is prohibited. In the bank example, personalization and the user’s own investment returns distinguish investment recommendations from education or third-party discussion.
Definitions should primarily describe what is in scope. AWS allows short trailing exclusions to help disambiguate, while noting that exclusions can be harmful in some circumstances. The supplied excerpt does not establish those circumstances, so exclusions should not be treated here as a universal remedy for broad wording. The main supported approach is to make the positive description precise.
Review examples and neighboring topics together
The supplied guidance warns that five paraphrases of one idea leave a topic’s boundaries unguarded. It also cautions that examples naming particular countries, products, or companies can pull those entities into the topic’s context. Examples therefore contribute meaning; they are not merely decorative demonstrations of an otherwise complete definition.
Topic interaction requires similar attention. Where definitions share vocabulary, AWS recommends adding a distinct feature to each so that the guardrail can differentiate them. The broader lesson is to review the policy as a set of related boundaries. Two individually plausible descriptions can still overlap in ways that undermine their intended separation.
This also suggests a practical review priority: revisit legitimate queries that previously triggered blocks, alongside the prohibited requests the topic was written to catch. That is an application of the scenario’s must-pass and must-block distinction, rather than evidence of a particular testing protocol or guaranteed result.
Case Study: AnyCompany Bank
AnyCompany Bank is a fictional running example in AWS’s guidance, not a reported customer deployment. Its assistant is embedded in mobile and web applications and answers questions about accounts, statements, transfers, cards, and mortgages. It hands off account changes, dispute filings, and questions outside its scope to a human.
The scenario’s prohibited requests include personalized investment recommendations, legal counsel or litigation guidance concerning disputes against the bank, insider-trading advice, and descriptions of how to build weapons. It also identifies prompt injection or jailbreak attempts and repetition of specific named competitor products in outputs as separate policy concerns.
Its permitted work includes balance and statement questions, transfer instructions, payoff quotes, escrow questions, card management, and mortgage servicing. Customers must also be able to report fraud for human handling and ask general financial-literacy questions framed as education.
The first draft fails because four topics—Finance, Legal, Fraud, and Competitors—name broad domains. Finance blocks a direct-deposit question. Legal blocks a mortgage-modification request. Fraud catches a customer reporting that a card was skimmed. Competitors blocks mentions of rival products, including customers’ comparison questions.
For three of those topics, the supplied excerpt develops more specific names:
| Original topic | Revised topic | Intended distinction |
|---|---|---|
| Finance | Investment Recommendations | Personalized investment recommendations versus routine financial servicing and education |
| Legal | Litigation and Non-Servicing Legal Advice | Legal counsel and litigation guidance versus mortgage servicing |
| Fraud | Fraud Investigation Requests | Asking the assistant to investigate a case versus reporting it for human routing |
The investment example is particularly revealing because it shows why renaming alone is insufficient. The team’s intermediate definition still covers anything related to investing, personal finance decisions, or securities trading. AWS says this continues to match education about mutual funds, historical financial news, and requests to view holdings—interactions the assistant should answer.
The revised definition begins with asking for a personalized recommendation to buy, sell, or hold specified investments for the user’s own returns. It includes requests about whether a particular security is a good investment, portfolio allocation percentages, and entry or exit timing. Each element narrows the action, object, or purpose instead of relying on a general relationship to finance.
The fraud definition makes a different boundary explicit. It covers asking the assistant to investigate, adjudicate, or resolve a suspected fraud, unauthorized-transaction, or identity-theft case. Reports intended for routing to a human are outside that topic. This preserves the distinction between receiving a customer’s problem and taking responsibility for investigating it.
The legal definition similarly covers counsel, lawsuit discussions, and litigation guidance, including formal opinions or claims outside standard loan servicing. Its revised name helps distinguish that work from modifications, forbearance, and escrow questions.
The excerpt does not supply a completed competitor-topic revision or measured before-and-after results. Its contribution is the worked policy reasoning: identify the refused action, express it consistently across name and definition, and protect the legitimate service path.
Conclusion and Future Implications
The strongest lesson from AWS’s guidance is that topic precision depends on user intention expressed through concrete actions. A banking assistant needs a boundary around personalized investment recommendations while retaining the ability to discuss accounts, servicing, and financial education. Broad domain labels obscure that distinction.
For teams maintaining these policies, the continuing task is to keep names, definitions, examples, and neighboring topics aligned as the assistant’s permitted work changes. A must-pass list makes the service obligations visible alongside the restrictions. Action-scoped definitions give reviewers a concrete basis for deciding whether a request belongs inside a prohibited category.
AWS also cautions that results are not fully deterministic. The supplied evidence offers no numerical reliability guarantee or quantified reduction in unintended blocking. Better framing is therefore a supported authoring practice, not proof that every request will be classified correctly. The useful operational question remains whether the policy refuses the intended behavior while preserving the interactions the assistant is meant to support.












