In 30 seconds

Good instructions tell an AI what its job is, how to behave, and when to stop and ask. But instructions guide; they do not guarantee. Anything that must always hold belongs in a guardrail, not only in the instructions.

Why this comes up

A lot of effort goes into wording instructions perfectly, in the belief that the right sentence will make the AI reliable. Clear instructions do help. But treating them as a guarantee leads to unsafe designs, because a model may not follow every instruction under unusual inputs. Knowing what instructions can and cannot do is the key to using them well. See what is a prompt.

What good instructions contain

  • The job. What the AI is for, in plain terms, matching its defined job and non-goals.
  • How to behave. Tone, format, and how to handle common situations.
  • The sources to use. Point it at the information it should rely on, and tell it to answer from that rather than from memory.
  • When to stop. Explicit guidance to say "I do not know" or to escalate when it is unsure or a case is outside its job.
  • Examples. A few good examples of the desired behavior often help more than long descriptions.

Write them plainly

Write instructions the way you would brief a capable new colleague: clear, specific, and concrete. Prefer plain statements and examples over long, abstract rules. If an instruction is ambiguous to you, it will be ambiguous to the model.

Do not rely on instructions for hard limits

This is the crucial point. "Never do X" in the instructions is a request the model usually honours, not a limit it cannot cross. If X must never happen, a wrong recipient, an over-limit payment, a prohibited action, enforce it with a guardrail in the surrounding system. Use instructions to guide behavior; use guardrails to enforce boundaries.

A small example

A team wrote careful instructions telling an agent never to share internal costs. Under an unusual question it did anyway. The reliable fix was not a stronger sentence but keeping internal costs out of what the agent could access. The instructions shaped everyday behavior; the access control enforced the boundary.

What this does not solve

Good instructions improve behavior; they do not make it deterministic or safe by themselves. They are one input to a probabilistic system. Guardrails, validation, testing and review are what turn well-instructed behavior into a dependable workflow.

Good enough to move on when

  • The instructions state the job, behavior, sources and stop conditions.
  • They are concrete, with examples where helpful.
  • Hard limits are enforced as guardrails, not just written in instructions.
  • You have tested how the AI behaves on unusual and prohibited inputs.