Vague instructions leave the recipient to define success. “Clean your room” might mean organizing everything properly or pushing clutter into a closet. AI behaves similarly by filling missing context with common assumptions. Better instructions explain the purpose, audience, relevant background, boundaries, examples, and standard for success while leaving appropriate room for judgment.
Tell a child to clean their room and you may return to find everything pushed into the closet.
The child can reasonably claim the room looks clean. The parent had a different outcome in mind but did not explain it.
AI instructions create the same problem. The system may satisfy the literal request while missing the user’s unstated expectation.
That does not mean every prompt needs to become a technical document. It means important expectations should not remain trapped in the user’s head.
Generic instructions produce generic patterns
Ask a child to draw a house without additional context.
The picture may contain a square, a triangle roof, a centered door, windows on either side, a round sun with rays, stick figures, and a few flowers. There is nothing wrong with the drawing. The child used a familiar pattern to fill an open request.
AI does the same.
Did you want a townhome or mansion? Modern or Victorian architecture? An urban street or rural landscape? Should anyone appear in front? What season, mood, perspective, or level of realism fits the purpose?
AI is unusually capable of producing something from very little. That capability can conceal the absence of a meaningful brief.
Explain the purpose
Begin with what you are trying to accomplish.
“Write an email about the new system” requires AI to decide why the email exists.
A clearer instruction might explain that the message should prepare employees for an upcoming transition, reduce anxiety, identify what is changing, and direct people to training without overstating what is already known.
Purpose helps the model prioritize information and tone.
Before prompting, answer:
- What outcome do I want?
- Who is the audience?
- What should they understand, feel, decide, or do?
- What would make the result useful?
Provide relevant context
Context may include the organization, audience, previous communication, constraints, source materials, risks, and current stage of work.
Do not provide sensitive information unless the tool and account have been approved for it. More context is not always better if the information is irrelevant or inappropriate to share.
Useful context tells AI what makes this situation different from the most generic version of the task.
For example:
- The audience includes frontline employees who do not work at computers.
- The initiative is still being evaluated, not yet approved.
- Managers will discuss the change before the email is sent.
- The organization uses plain language and avoids exaggerated claims.
- Employees are concerned about whether automation will eliminate jobs.
Those details shape the response substantially.
Define boundaries
Good instructions clarify what should not happen.
Boundaries might include:
- Do not invent statistics or quotations.
- Do not change the physical structure shown in the source image.
- Do not name employees.
- Avoid technical jargon.
- Do not imply a decision has been finalized.
- Keep the response below 500 words.
- Identify uncertainty rather than filling gaps.
- Use only the supplied source materials.
Constraints reduce avoidable variation and make review easier.
Show examples and standards
Examples help communicate a desired pattern, especially when tone or format is difficult to describe.
Provide a previous article, email, report, or design that represents the organization well. Explain what should be preserved, such as its direct opening, short sections, practical examples, or restrained call to action.
Also state how the result will be judged.
Official prompting guidance from Anthropic similarly emphasizes clear direction, relevant context, examples, and control of the requested format.
A model can respond more effectively when it knows the destination.
Ask AI to clarify
The user does not need to anticipate everything in one giant prompt.
Ask AI what additional information would materially improve the result. It may identify missing audience information, conflicting constraints, unclear measurements, or decisions that need human judgment.
This turns prompting into a conversation rather than a test of whether the user can write one perfect instruction.
The user should still decide which questions matter. AI can ask for context, but it cannot determine the organization’s purpose or values.
Leave room for useful judgment
Detailed instruction should not become unnecessary micromanagement.
If the task involves brainstorming, exploration, or creative alternatives, excessive constraints may prevent useful ideas. Define the important boundaries and allow variation elsewhere.
The right level of instruction depends on the consequence. A low-stakes list of ideas can remain open. A policy, public statement, financial analysis, or safety procedure needs greater specificity and verification.
Inspect the closet
The final step is review.
Did the response meet the real expectation, or did it merely make the surface look clean? What assumptions are hiding behind the answer? What was omitted? What needs to be corrected, sourced, or adapted?
AI will confidently fill the space left by vague instructions. Human judgment must decide whether what filled that space belongs there.
Before blaming the model for getting the wrong answer, examine the instruction. You may have said “clean the room” while imagining a result you never described.
Put the idea to work
### What could this make possible for your organization?
Bring Foster + Foundry the challenge, opportunity, or question. We will help clarify a practical next step.
