AI Rules: Set Clear Boundaries for Your Assistants

AI rules are the instructions and boundaries that govern how an assistant works: what information it may use, which actions it may take, and when it must stop for a person. The phrase can also mean company policies or legal requirements, which need different owners and methods of enforcement.

Start by separating those meanings. A sentence in a prompt can guide a draft, but it cannot replace access controls or settle a legal question. fenbs keeps approved working rules with the project so connected assistants can read the same instructions before continuing the work.

Which kinds of rules are you setting?

Working instructions explain how to complete a particular job. Team policies establish what people may use AI for across the business. Legal requirements come from the authorities and obligations that apply to your activity. Keep their sources visible so an assistant cannot mistake a convenient preference for permission to ignore a requirement.

The NIST AI Risk Management Framework↗ (opens in a new tab) offers voluntary guidance for managing AI risks. It is not a law. For wider questions about ownership and review, see the AI governance guides.

What should a useful working rule contain?

Write each rule around a decision the assistant will actually face. Name the work it covers, the allowed action, the boundary, and the person to ask when the boundary is reached. Add a reason when it helps someone recognize a similar situation.

• Scope: customer-facing copy for this project, rather than every sentence anyone writes.
• Instruction: use only claims supported by the approved source list.
• Stop condition: if a claim lacks support, flag it instead of inventing evidence.
• Owner: the marketing lead approves a new claim before publication.
• Reason: drafts must not turn an assumption into a promise to customers.

Avoid instructions such as “be responsible” on their own. Two people can agree with that phrase and still disagree about sending a document to an outside service. Concrete boundaries are easier to follow, test, and explain to a new teammate.

Keep the first set short enough to read before work starts. The guide to using AI at work helps identify a small workflow where these boundaries can be tested without granting broad authority.

Which boundaries deserve explicit instructions?

Begin with data, actions, evidence, and approval. These are common points where a helpful suggestion can become an unwanted change. State what the assistant should do when it is uncertain, rather than expecting it to infer the safest next step.

• Data: use approved project material; stop before sharing restricted information with another service.
• Actions: prepare drafts and proposed changes; obtain the named approval before sending or publishing.
• Evidence: distinguish checked facts from assumptions and link the sources supporting material claims.
• Completion: report what was tested and what remains unchecked before calling a task finished.
• Conflict: explain incompatible instructions and ask the responsible person to resolve them.

Say a small architecture studio uses AI to draft a project update. A rule can allow summaries of approved milestones while requiring review before promising a completion date. The assistant should flag an uncertain date; a project manager decides what the client can be told.

That approval belongs in the shared record. Decisions and rules in fenbs preserve what a person agreed and show continuing rules first in the context connected assistants read.

Can a prompt enforce a company policy?

A prompt provides instructions. Enforcement requires controls at the place where the action happens, plus people checking that the controls fit the risk. If an assistant should not send email, removing its sending access is stronger than relying only on a sentence asking it to wait.

Use approved accounts, suitable permissions, and review before consequential actions. Test the boundary with an ordinary request that would exceed it. If the assistant can still perform the forbidden action, improve the actual access control rather than adding louder language to the prompt.

Tool-specific instruction files can help deliver local conventions; Cursor project rules cover that narrower setup. For the neighboring questions, see Windsurf AI rules, the wider governance process, and enterprise rollout rules.

How do legal requirements differ from team preferences?

Internal approval does not establish that a use is lawful. The FTC statement on existing laws and AI↗ (opens in a new tab) explains that existing legal protections still apply to automated systems. Which requirements apply depends on the work, information, people affected, and jurisdiction.

Have a qualified reviewer identify the requirements for a consequential use, then translate the approved conclusion into working instructions. Keep a reference to the underlying advice and its scope. An assistant should not decide that a new hiring, financial, or health-related workflow is permitted because a different project received approval.

How do you keep rules current across assistants?

Give every continuing rule one current home, an owner, and a clear scope. Store reusable prompt wording separately from the decision that authorizes it. A prompt can change for clarity while the underlying boundary stays the same.

The NIST AI RMF Playbook↗ (opens in a new tab) provides voluntary suggested actions that organizations can select for their circumstances. Use it as a source of review questions, then have people choose the rules your work actually needs.

When a rule changes, record the new decision and which earlier instruction it replaces. Review it after a new tool connection, a significant workflow change, or a mistake. A shared AI brain is useful when it preserves that current work record without making old chats the authority.

How to do this with fenbs

1. Create a free account, free for one person with no credit card, and add the project whose working rules you want to organize.
2. Record the agreed boundaries in Decisions and rules. Identify the human decider and the work each rule covers; leave unresolved choices open for a person.
3. Save reusable instructions in Skills and prompts, keeping the approved rule distinct from example wording.
4. Connect a supported assistant using the integration guide. Ask it to read the project context before working and explain the approval boundary for its next action.
5. Save progress at the end, including unchecked work. If a problem reveals an unclear rule, record the lesson and ask the responsible person whether the rule should change.

fenbs is useful for keeping approved instructions available across connected assistants. It does not enforce another app’s settings, provide legal advice, or replace security controls; use dedicated systems when those are the job you need done.