Skip to content
All frameworks

Template · Chapters 2 & 5

Constraint System Template

A structured document that defines what your AI system can and cannot do, across six dimensions. Write it before the first production cycle runs. The version in the document is the one the system runs.

What it is

The brand contract tells the system how to sound. The constraint system tells the system what it is allowed to do. Both are design artifacts. Both belong in writing before the product ships.

Most teams make constraint decisions in isolation: one engineer sets a confidence threshold, another decides what the agent escalates, a third defines what it remembers. The result is a system whose constraints are inconsistent, undocumented, and hard to audit when something goes wrong.

The constraint system template gathers all six constraint dimensions into one document. Fill it in together. Keep it current. When a constraint changes, the document changes first.

The six dimensions

Each dimension is a separate design decision. Each decision needs an owner.

01

Agent Role Definition

Role name

What this agent is called in the product and in the codebase.

Scope

What this agent handles. Be specific. Broad scope statements are not constraints.

Out of scope

What this agent explicitly does not handle. Name it before launch, not after.

02

Behavioral Tokens

Confidence threshold for autonomous action

Below this score, the agent does not act without approval. Example: 0.80.

Confidence threshold for draft vs. act

Below this score, the agent drafts a response rather than sending it. Example: 0.65.

Behavior under uncertainty

What the agent does when confidence is below either threshold. Does it ask? Escalate? Draft?

Rationale

Why these thresholds. Reference the data that supports them. Revisit when the data changes.

03

Permission Levels

Read only

What the agent can see but cannot act on.

Suggest only

What the agent can recommend but not execute.

Draft and propose

What the agent can prepare for human review before sending.

Execute with approval

What the agent can do after a human confirms.

Execute autonomously

What the agent can do without asking. These actions must be reversible or low-stakes.

04

Memory Rules

What the agent remembers

Across sessions, across contexts. Name it explicitly.

What the agent forgets

After each session, after a defined period, or when the user requests it.

What the agent never stores

Sensitive data, health information, payment details. List it. This is a privacy specification.

Correction mechanism

How does a user correct or remove stored context? If you cannot answer this, the memory design is not done.

05

Refusal Register

Capability refusals

What the system cannot do. Legible response for each.

Policy refusals

What the system will not do. Legible response for each.

Confidence refusals

What the system declines when it is not certain enough to act.

Scope refusals

What the system redirects because it is outside the product's intended use.

A full refusal register template is at artofuserexperience.com/resources/refusal-register.

06

Human-in-the-Loop Position

High-stakes actions

Where is the human for actions that are irreversible or consequential? Early gate, continuous monitor, or late review?

Low-stakes reversible actions

Late review is usually right here. Document the decision.

Handoff condition

The specific trigger that moves control from agent to human. See artofuserexperience.com/resources/handoff-condition-template.

Escalation owner

Who receives the handoff? "A human" is not an answer.

How to use it

Fill in all six dimensions before the first production cycle. A blank field is a decision that has been left to chance. The PM owns the constraint system. Engineering, design, and legal contribute to it.

Review it every sprint that touches behavior the document defines. The version in the document is the one the system runs. If the system is behaving differently from what the document says, one of them is wrong. Fix the one that is wrong.

For solo practitioners: fill in what you know. Mark what you do not know. The document's gaps are your design backlog.

Related templates

From the book

Chapters 2 and 5 cover the Human and Agent Experience System and how agents take action on behalf of users.