Most people learn to use AI by writing instructions: write an email, summarize this document, analyze this spreadsheet, create a blog post about this topic.
That works well for straightforward tasks. You know what you want the AI to do, you give it instructions, and it gives you an answer. As the work gets more complicated, though, there is a better way to approach prompting.
Instead of focusing entirely on the individual task, start by defining the outcome.
- What are you actually trying to accomplish?
- What would a good result look like?
- Does the AI have enough information to get there?
- How should it know when it has done enough?
That shift changes the way you write a prompt and, in some cases, changes how the AI approaches the work. Two useful methods are loops and goals, supported by guardrails that keep the work within useful boundaries.
For more complex tasks, another simple addition can help before any work begins: giving the AI permission to ask questions when important context is missing.
- Loops give the AI a process for reviewing and improving its work.
- Goals define an outcome while allowing more flexibility in how to reach it.
- Guardrails establish the limits, rules, and conditions the AI should follow along the way.
The Limits of a One-Shot Prompt
A traditional prompt might look something like this:
Write a 500-word article about workflow automation for marketing managers. Include examples and make the tone conversational.
There is nothing wrong with that prompt. The AI has a task, some context, and a few instructions. For many jobs, that is enough.
The limitation is that the AI gets one pass at the problem. You may get a strong result, or you may get something that technically follows the instructions while still missing what you wanted.
That usually leads to another prompt:
Make it less generic.
Then maybe:
Add more practical examples and make it sound less like AI wrote it.
At that point, you have created a loop manually. You are reviewing the result, identifying what is weak, giving the AI another instruction, and repeating the process.
A way to improve the process is simple: build the review into the original prompt.
Method One: Build a Loop Into the Prompt
A loop tells the AI to create something, evaluate the result against specific criteria, make improvements, and check the work again.
A simple loop might look like this:
GOAL:
Write a useful LinkedIn post about workflow automation.
PROCESS:
1. Write the first version.
2. Review it for clarity, specificity, audience relevance, and unnecessary jargon.
3. Identify the weakest area.
4. Revise the post.
5. Review the revised version.
GUARDRAILS:
Do not introduce unsupported claims.
Do not exceed two revision passes.
STOP WHEN:
The criteria are satisfied or two revision passes have been completed.
OUTPUT:
Return the final version.
The evaluation step is what makes the loop useful. The AI has something specific to check before it considers the work complete.
That can improve tasks where the first answer is really a draft. Writing is an obvious example, but the same approach can work for analysis, code review, structured data, planning, and other work with reasonably clear quality criteria.
Why Loops Work
When I write something myself, I rarely expect the first version to be the finished version. I write it, read it back, see what’s missing, make changes, and eventually reach a point where another revision is unlikely to materially improve the result.
AI can follow a similar process if we tell it how.
Consider a website review. A basic prompt might say:
Review this page and recommend ways to improve it.
A loop-based prompt could be more specific:
Review this page and identify the changes that would most improve clarity for a potential customer.
After creating your recommendations:
1. Check whether each recommendation addresses a real problem on the page.
2. Remove generic recommendations that could apply to almost any website.
3. Make each remaining recommendation specific enough for someone to implement.
4. Review the final list against the original page.
GUARDRAILS:
Base recommendations only on what is actually present on the page.
Do not create recommendations simply to increase the number of findings.
Stop after two review passes.
The request is still a single prompt. What changes is the process inside it. The AI now has a method for deciding whether the first answer is good enough.
Where Loops Help Most
Loops work especially well when you already understand the process required to produce a good result. You know what should be checked, what quality looks like, and what the AI should do when something falls short.
Loops can also reduce some of the back-and-forth that happens during normal prompting. Instead of waiting for the first response to reveal obvious problems, you ask the AI to look for those problems before it finishes.
There are limits. More passes do not automatically create better work. An AI can revise something that was already good and make it worse, or spend extra effort fixing details that have little impact on the final result.
That is why I would avoid prompts like:
Keep improving this until it is perfect.
There is no useful definition of perfect.
A better loop has clear criteria, guardrails, and a limit. For many tasks, one or two review passes may be enough. If the AI still cannot satisfy the requirements after several attempts, the problem may be missing context, weak source material, or a question that requires human judgment.
Method Two: Give the AI a Goal
The second approach is more flexible.
Instead of describing every step the AI should take, define the outcome you want it to achieve.
For example:
/goal
Determine why organic leads declined last month and provide evidence-supported explanations that a marketing manager can act on.
I am using /goal here as shorthand for a goal-oriented instruction. The exact syntax depends on the AI system you are using. Some environments may support a specific goal command, while others simply need the goal written into the prompt.
The important idea is that the desired outcome stays clear while the path toward it can change. That becomes valuable when you cannot predict every useful step in advance.
Goals Work Well When the Path is Uncertain
Imagine you are investigating a decline in organic leads.
You could write a detailed prompt:
Compare traffic month over month. Identify pages with declining traffic. Check search rankings for those pages. Review conversion rates. Summarize the findings.
That gives the AI a sensible process, but it also assumes you already know where the problem is likely to be.
Maybe traffic is stable and the real issue is conversion rate. Maybe a tracking change affected the numbers. Maybe one high-value landing page disappeared from the site.
A goal-oriented prompt can leave room for those possibilities:
GOAL:
Determine the most likely causes of the decline in organic leads.
CONTEXT:
You have access to GA4 and Search Console data.
GUARDRAILS:
Base findings on available evidence.
Separate confirmed findings from hypotheses.
Do not assume ranking loss is the cause.
Identify missing information when it could materially change the conclusion.
SUCCESS LOOKS LIKE:
A clear explanation of the major factors contributing to the decline and practical next steps.
The AI knows what it is trying to accomplish without being locked into one exact path.
That distinction becomes more important as AI systems gain access to tools, files, analytics platforms, email, search, and other sources of information. The AI may be able to decide what information it needs based on what it has already learned.
Where Goals Help Most
Goal-based prompting is useful for research, troubleshooting, analysis, planning, and other work where each finding may change what should happen next.
It also reduces the burden of predicting the entire process ahead of time. Sometimes we know the result we need without knowing the exact sequence required to get there. A goal lets us define that result clearly while giving the AI room to determine the method.
This starts to look more like delegation. If I ask someone on my team to investigate a problem, I may explain the issue, provide context, define what I need to know, and give them boundaries. I would not necessarily prescribe every individual step.
Goal-based AI can work in a similar way.
Sometimes the Best First Step is a Question
Defining a good outcome does not mean you always have enough information to reach it.
For more complex work, one of the most useful instructions you can add is permission for the AI to ask questions before it begins.
This is especially helpful when missing context could materially change the answer.
For example:
BEFORE STARTING:
Ask up to five questions that would materially improve the result.
Focus on missing context, the desired outcome, available information, and important constraints.
If enough information is already available, proceed without asking unnecessary questions.
This can prevent the AI from filling in gaps with assumptions.
Suppose the goal is:
GOAL:
Help determine why website leads have declined.
BEFORE STARTING:
Ask any questions needed to understand the timeframe, available data, recent site changes, and how leads are being measured.
GUARDRAILS:
Ask only questions that could materially change the analysis.
Limit clarification to five questions.
Now the AI has another option beyond immediately producing an answer. It can first determine whether the problem is defined well enough to solve.
That matters because we often know things the AI does not. There may have been a website launch, a tracking change, a change in media spend, or a shift in how leads are counted. A few focused questions can surface that information before the AI starts working in the wrong direction.
I would still put limits around this. An instruction such as “ask me whatever you need to know” can result in unnecessary questioning. A better approach is to ask only for information that could meaningfully affect the outcome.
Goals and Loops Still Need Guardrails
The more freedom you give the AI, the more important guardrails become.
Guardrails define the boundaries of the work. They can include constraints on what the AI may do, which sources it can use, how many times it can iterate, when it should ask questions, when it should stop, and when a human should take over.
For example:
GOAL:
Identify opportunities to improve lead quality from our paid search campaigns.
CONTEXT:
Focus on the last 90 days of campaign and conversion data.
BEFORE STARTING:
Ask questions only if information is missing that could materially change the analysis.
GUARDRAILS:
Use only the supplied campaign and conversion data.
Do not recommend increasing budget unless the data supports it.
Clearly distinguish observations from recommendations.
Do not make changes automatically.
Flag any recommendation that requires human approval.
SUCCESS LOOKS LIKE:
A set of changes the paid media team can evaluate and implement.
The AI still has room to explore the problem, but that freedom exists within a defined set of boundaries.
Guardrails are useful with loops, too. A loop without limits can keep revising long after the additional work stops being valuable. A goal without boundaries can become too broad or wander into areas that were never part of the assignment.
A useful way to think about the relationship is:
Goal = the outcome
Loop = the process
Guardrails = the boundaries
Questions sit slightly differently. They help fill in missing context before or during the process when the AI does not yet have enough information to pursue the goal effectively.
From Prompt to Loop to Goal
The graphic above shows the relationship between the three approaches and how guardrails help define the boundaries around more flexible AI workflows.
Loop or Goal?
The easiest way to decide between them is to think about how well you understand the path to the result.
If you already know the process, a loop can work very well. That could include writing a draft and reviewing it against brand guidelines, validating structured data, reviewing code against known requirements, or checking that a document covers every required topic.
A goal becomes more useful when the process may need to change based on what the AI discovers. Research, technical troubleshooting, analytics investigations, document review, and preparation work often fit this category.
In either case, consider whether the AI has enough information before it begins. If an important detail could substantially change the work, giving it permission to ask a focused question can improve the result before the loop or goal process even starts.
Guardrails matter throughout. They keep the AI from taking unnecessary actions, continuing for too long, relying on the wrong information, or making decisions that should remain with a person.
The methods can also work together. In many cases, combining them produces a stronger result.
Combining a Goal, Questions, a Loop, and Guardrails
A goal can define what needs to be accomplished while a loop gives the AI a controlled way to keep working toward it. Questions help fill important gaps before the work starts, and guardrails keep the process within acceptable limits.
For example:
GOAL:
Produce the strongest recommendation for improving this landing page.
CONTEXT:
The page is designed to generate qualified B2B leads.
BEFORE STARTING:
If critical information about the audience, offer, or conversion goal is missing, ask up to three questions before beginning.
Otherwise, continue directly to the review.
CRITERIA:
Recommendations should be specific, supported by evidence from the page, and realistic to implement.
PROCESS:
1. Review the page.
2. Identify the most important issue affecting the goal.
3. Develop a recommendation.
4. Evaluate the recommendation against the criteria.
5. Revise it if the criteria are not satisfied.
GUARDRAILS:
Base recommendations only on the supplied page and provided context.
Do not invent missing performance data.
Do not recommend a complete redesign unless the evidence clearly supports it.
Stop after three passes.
STOP WHEN:
The recommendation satisfies the criteria or three passes have been completed.
OUTPUT:
Provide the final recommendation and the evidence supporting it.
This gives the AI a destination, enough context to determine whether it can begin, a process for doing the work, and boundaries around how far it can go.
For more complex AI systems, the same ideas can extend beyond a single prompt. An agent might ask for missing information, inspect available data, choose an action, evaluate what happened, and decide whether another action is needed. Guardrails can control which tools are available, what data can be accessed, and where human approval is required.
The underlying concept remains useful even when the technology behind it becomes more sophisticated.
A Simple Framework for Better Prompts
You do not need to turn every prompt into a complicated framework. For a simple request, a simple prompt is usually enough.
When the task becomes more important or less predictable, I find it useful to think through a few questions before writing the prompt:
- What result do I actually need?
- What context does the AI need to understand?
- Is anything important missing that the AI should ask about first?
- What makes the result acceptable?
- Is there a reliable process the AI should follow?
- Would a review pass improve the result?
- What boundaries should the AI stay within?
- How should the AI know when the work is complete?
A reusable prompt could look like this:
GOAL:
[What outcome do I need?]
CONTEXT:
[What does the AI need to understand?]
BEFORE STARTING:
[What should the AI clarify if important information is missing?]
CRITERIA:
[What makes the result successful?]
PROCESS:
[What method should the AI follow, if one is required?]
GUARDRAILS:
[What boundaries, limits, sources, permissions, or rules should it follow?]
STOP WHEN:
[How should it determine that the task is complete?]
OUTPUT:
[What should the final result contain?]
You will not need every section every time. The value of the structure increases as the task becomes more complex.
Prompting is Becoming More About Defining the Work
For a long time, prompt writing has focused heavily on finding the right words to get the right answer from AI. That still matters, but more capable AI systems make it useful to think beyond the wording of a single request.
The bigger questions are about the work itself.
- What outcome are we trying to reach?
- Does the AI have enough context to begin?
- What information should influence the result?
- How much freedom should it have?
- What boundaries should apply?
- What does completion actually look like?
Loops, goals, questions, and guardrails are practical ways to start answering those questions.
Loops help when the process is understood and the AI can improve its work against defined criteria. Goals are useful when the path may change based on what the AI discovers. Questions can fill important gaps before the AI heads in the wrong direction, while guardrails define the limits around the work.
As AI starts working across more of our files, systems, and everyday processes, I expect this way of prompting to become increasingly important.
The quality of the result depends heavily on how clearly we define the outcome, the context, and the boundaries around how AI gets there.
Still not sure where to start?
That’s ok, AI continues to move fast and change quickly. Arc intermedia provides AI Consulting as well as many other AI Solutions. Contact us to speak with an expert and find out how we help take your AI usage to the next level.
