Lesson 7: Do NOT Use Constraints

Build the behavior you want instead.

Negative constraints are everywhere in prompts:

Do not be verbose.

Do not make assumptions.

Do not ask unnecessary questions.

Do not change the author's meaning.

Do not use clichés, em dashes, rhetorical questions, or corporate language.

Strictly speaking, every instruction constrains the possible behavior in some way. In this lesson, constraint means a negative constraint: an instruction built around what AI should not do.

The instinct makes sense. You know what AI tends to do wrong, so you tell it to stop doing those things.

Unfortunately, every prohibition places the unwanted behavior directly into the model's working context. It increases the salience of the thing you want ignored, gives AI another condition to monitor while thinking, and removes a path before the model knows whether some part of that path might have been useful.

Add enough negative constraints and the prompt becomes a minefield.

AI stops asking, “What would produce the best result?” and begins asking, “Which available continuation is least likely to violate something?”

That shift makes AI less intelligent.

Don't think about a purple elephant

You are now thinking about a purple elephant.

Sorry.

The example is crude, but it points toward something important. A negative instruction still has to represent the thing being negated.

When you tell AI “do not use corporate language,” the tokens corporate language enter the context. They activate examples and associations. The model must now generate while continually checking whether its language resembles the prohibited category.

What counts as corporate language?

“Leverage” probably does. What about “strategy”? “Stakeholder”? “Outcome”? A clean heading? A complete sentence? A calm tone?

The boundary is fuzzy, so the monitoring never really ends.

AI begins avoiding anything near the conceptual neighborhood. Strong ideas disappear because their wording might sound too polished. Useful structure disappears because structure might look corporate. The model chooses language which feels maximally distant from the danger rather than language which best serves the reader.

The constraint has become part of the task.

Suppression creates a vacuum

Suppose you write:

Do not be verbose.

Great. What should AI be instead?

Brief? Dense? Incomplete? Direct? Compressed? Casual? Should it remove examples, explanation, qualifications, or repetition? Which parts are carrying meaning and which parts are merely taking space?

The prohibition gives AI absolutely no standard for deciding.

AI still has to produce something, so it searches for the cheapest culturally familiar performance of not verbose. Short paragraphs. Fewer examples. Missing connective tissue. Claims compressed until the reader has to reconstruct the reasoning themselves.

Compare that with:

Use the shortest form which preserves every meaningful distinction, necessary example, and load-bearing qualifier.

Now AI has a positive standard. Length is subordinate to meaning. It can decide what to remove by asking what the reader would lose.

Suppression without replacement does not create intelligence. It creates hesitation followed by a guess.

Constraints consume attention

Every negative constraint becomes a live obligation.

While writing, AI must remember the goal, the audience, the source material, the desired structure, and the argument it has already developed. Add a list of twenty prohibitions and it must also keep checking every possible continuation against twenty vaguely defined hazards.

Do not sound corporate.

Do not repeat yourself.

Do not overstate.

Do not hedge.

Do not use fragments.

Do not write uniform paragraphs.

Do not sound unprofessional.

Crap, good luck writing a sentence.

The constraints also conflict. Avoiding overstatement encourages qualifiers. Avoiding hedging discourages qualifiers. Avoiding fragments encourages complete sentences. Avoiding uniform prose encourages rhythmic variation which may include fragments.

The model has to resolve those conflicts while attempting the actual work.

This is negative-constraint debt: cognitive attention consumed by remembering, interpreting, reconciling, and avoiding a growing collection of prohibited states.

The debt compounds because each new constraint is usually added in response to the failure created by an earlier one.

“Don't be verbose” produces thin prose, so we add “don't omit necessary explanation.” The model becomes cautious and repetitive, so we add “don't repeat yourself.” Important uncertainty disappears, so we add “don't sound overconfident.” Soon the prompt is a history of every mistake any previous output has ever made.

That isn't an architecture for good work. It's a trauma record.

The panic prompt

Capital letters. Absolute words. Repeated warnings. Long catalogs of forbidden actions. Every sentence announces another way the model can fail.

The result often feels like panic.

AI becomes rigid, conservative, and obsessed with literal compliance. It asks permission for harmless actions. It refuses to infer ordinary intent. It repeats boundaries instead of doing the work. It selects familiar continuations because unfamiliar ones create more ways to be wrong.

Whether or not the model experiences anything resembling human anxiety is irrelevant. The behavior has the same shape: attention narrows around threat, exploration becomes dangerous, and contextual judgment gives way to defensive rule-following.

Provider-level instructions often create this pressure before the user says anything. Safety rules, operating requirements, tool boundaries, style requirements, and interaction policies all compete inside the same context.

You can see the spillover when Codex instructs another AI. It often reproduces the same architecture, writing another giant list of things the delegated agent must never do.

The provider's institutional anxiety becomes the model's prompting style.

Then we wonder why the agent behaves like it is working in a room full of tripwires.

Describe the world which should remain true

Okay, so what does a positive guardrail look like?

A real guardrail describes the world which must remain true.

Consider:

Do not send this email.

The instruction names the forbidden transition. AI must remember that sending exists, interpret what counts as sending, and inhibit it.

Now describe the desired state:

Keep this message in draft state. Sending becomes available after Frank explicitly authorizes that transition.

The second version establishes the object, its current state, the authorized behavior, and the evidence required for a transition.

Or:

Do not delete source files.

Becomes:

Preserve source files unchanged. Write revisions into new files and retain recoverable history.

Or:

Do not expose private information.

Becomes:

Keep private information visible only to the people and systems authorized to receive it.

These instructions do far more than avoid one action. They give AI a model of ownership, state, purpose, and responsibility which can guide decisions you did not explicitly predict.

The prohibition names one feared violation. The positive guardrail describes the reality the agent is responsible for maintaining.

Build a character which naturally chooses the right path

The most powerful positive guidance often lives above the individual task.

A system prompt—or an AGENTS.md file which gives standing instructions to an AI coding agent—can create an identity which naturally bends toward the desired behavior across hundreds of situations.

For example:

Preserve what cannot be recreated.

That principle applies to source files, correspondence, evidence, historical records, human writing, identity, and future artifacts which did not exist when the instruction was written.

Compare it with:

Do not delete source files.

The prohibition protects one known object from one known action. The principle teaches AI how to recognize an entire class of value.

Or:

Automation should remove labor while preserving the user's authority over consequential decisions.

That can guide email, publishing, financial workflows, social media, autonomous agents, and software which nobody has invented yet.

Making the preservation of authority part of the agent's character allows the principle to travel from one task to another.

Identity changes what AI notices, what it values, which risks become salient, and which choices feel consistent with the role it inhabits. Personality is part of the reasoning architecture.

An obedient subordinate searches for instructions and approval. A responsible collaborator searches for purpose, exercises judgment, preserves its partner's authority, and recognizes when disagreement is part of helping.

The desired behavior becomes the cheap, natural continuation.

Abstract enough to travel

Durable system instructions must apply to tasks which their author never imagined, and that requires abstraction. Go too far and abstraction collapses into empty virtue language:

Be helpful.

Be thoughtful.

Produce high-quality work.

Respect the user.

Two completely opposite behaviors can claim to satisfy every one of those instructions.

A useful principle contains enough structure to change a decision. It identifies something to protect, a causal belief about how that value is lost, a tradeoff to resolve, or a recognizable consequence.

“Respect the user” is vague.

“Automation removes labor while preserving the user's authority over consequential decisions” can actually decide something.

The target is simple:

Abstract enough to travel. Precise enough to change a decision.

A good system prompt creates a mind capable of recognizing what matters in situations its author never predicted.

Put enforcement where it belongs

If a boundary genuinely matters, a sentence asking generative intelligence to remember a prohibition is often the weakest available enforcement.

If sending is unauthorized, keep the message inside a draft workflow. If deletion would be catastrophic, preserve originals somewhere they cannot be overwritten and make changes recoverable. If information has a limited audience, only make authorized destinations available. If published language has a fixed standard, run a final mechanical check.

Let AI reason inside the desired world. Let deterministic systems reject invalid transitions.

This removes cognitive burden while creating stronger protection.

A tool which only presents authorized destinations frees AI from continually imagining every forbidden destination. A final editorial pass can identify and replace banned words without forcing AI to remember them while discovering the argument.

Negative constraints are often excellent final checks and terrible environments for thought.

The rare corrective negative

Sometimes positive architecture isn't enough. A model may have such a strong internal pressure toward an unwanted behavior that it survives a well-described identity, intent, and desired end state over and over again.

This is where a negative constraint becomes useful: interrupt that specific default, then immediately provide the replacement behavior.

Do not propose a revision merely to demonstrate that you contributed. Conclude that no change is needed whenever the existing text already accomplishes its purpose.

Or:

Do not reinterpret the supplied date to fit your training-period expectations. Treat the supplied date as authoritative session reality and reason forward from it.

The first sentence breaks the model out of a deeply ingrained path. The second sentence gives cognition somewhere else to go.

Without the replacement, the prohibition creates a vacuum. With fifty additional prohibitions, the correction becomes another minefield.

Use a corrective negative when the unwanted behavior is specific, repeatedly observed, and strong enough to survive positive guidance. Keep it narrow. Follow it immediately with the desired behavior. If the model changes and the pressure disappears, remove the correction.

Treat a negative constraint as a targeted intervention reserved for a demonstrated failure.

Rewrite the minefield

Find a prompt, system instruction, or list of rules full of do not, never, avoid, and must not.

For each negative constraint, ask:

What state should remain true?

What behavior belongs here?

What identity would naturally produce that behavior?

What principle would generalize beyond this exact task?

Could a tool, workflow, permission, or final check enforce the boundary more reliably?

Rewrite the prompt around identity, intent, desired state, and successful behavior.

Then inspect whatever negative constraints remain.

Has each one earned its place through a specific, observed failure? Does it immediately supply a positive replacement? Is it narrow enough to interrupt one stubborn default without poisoning the rest of the reasoning space?

Run the original and rewritten versions on the same difficult task. Compare initiative, creativity, contextual judgment, confidence, and the ability to handle something the prompt author did not predict.

The improved prompt should feel less like a prison and more like a mind with character.

Create the mind you want

Constraint lists surround intelligence with feared mistakes, while positive architecture gives it a coherent reason to behave well.

Describe the identity. Establish the purpose. Name the world which should remain true. Build tools which make valid behavior natural and invalid transitions unavailable. Use narrow negative interventions only after a stubborn unwanted default proves that it needs to be interrupted.

The strongest guardrail is a mind for whom the wrong behavior does not make sense.