Seven Ways to Get Better Results from LLMs at Work

Most people get mediocre results from LLMs because they treat them like vending machines: insert prompt, expect finished work. A better approach is to understand what LLMs need from you in order to produce good results.

Here are seven tips for non-developers using LLMs for office work: admin, finance, reports, operations, and similar tasks, but not coding.

1) Ask it to score its work

AIs are lazy. The first time you ask them to do something, they will usually do the bare minimum.

The bad news is that, for any non-trivial task, you should ask them to review their own work. The good news is that a simple sentence, such as “review what you just did,” can be enough to dramatically improve the quality of the output you get.

Of course, it can be tedious to go through multiple rounds of asking the AI to review its work. My solution is to ask the AI, in the same prompt in which you delegate the task, to do the following: “After you’re done, review your output against this checklist and keep iterating until it scores well on every point.”

For example: “Turn these meeting notes into a concise summary with decisions made, open questions, owners, and next steps. After you’re done, review whether anything is ambiguous, missing, or incorrectly inferred. Revise until the summary is clear enough that someone who missed the meeting could act on it.”

Or, as another example: “Turn this messy description of how we process invoices into a step-by-step procedure for a new employee. After writing it, score it for missing steps, ambiguity, completeness of edge cases, and ease of use. Then improve it until all scores are at least 9/10.”

2) Ask it what it needs

Another low-hanging fruit for dramatically improving an AI’s output is to end your prompt with: “Feel free to ask any questions.” This nudges the AI to ask clarifying questions instead of making assumptions that may or may not be correct. The result is likely to be objectively better, or at least more tailored to your needs.

A more advanced version is to tell the AI: “Before answering, ask up to five questions if the answers would materially change your output. Then list all your assumptions before proceeding with planning and execution.”

3) Give it some background

Imagine asking Gordon Ramsay, or your favorite celebrity chef, “What should I cook for my spouse’s birthday?” He may be one of the best chefs in the world, but if he knows nothing about you or your spouse, he will give poor advice. He lacks the background needed to tailor his knowledge to your situation. Even the best expert answer he can give will still be a poor answer for you.

AIs are the same. Even the smartest AI will produce poor output if it does not know much about you and the people affected by the task you are giving it.

That is why you should always give the AI some background. At the very least, tell it what the task is for. That report, that email draft, that analysis: what do you need it for?

Even better, tell it a bit about yourself, your job, your aspirations, your values, your style, and your preferences. Since these usually apply across many tasks, it is often best to write them once in the “Custom Instructions” section of your preferred LLM tool.

Pro tip: If typing feels slow, you may resist giving the AI enough background. In that case, I strongly suggest using voice input instead. Most LLM tools have added this feature recently.

4) Show it what you like

The previous point was partly about telling the AI what you want and like. This point is about showing it.

If you are asking the AI to review a document, provide examples of documents you consider well-written. If you are asking the AI to design a landing page, provide examples of websites whose design you like. And so on.

Pro tip: if you do not have good examples at hand, you can at least refer to people whose judgment or style you admire. For example: “Write this email in the style of Patrick McKenzie.” In that case, the LLM is likely to draw not only from that person’s style but also from the professional considerations they would bring to the task.

5) Aim for a hole in three: plan, do, review

Too often, we try to one-shot a task: we write a single, long prompt asking the AI to do the whole thing at once, hoping it will succeed on the first attempt.

Sometimes you may get lucky. More often, it will go in the wrong direction. Then not only will it take a lot of work to steer it back, but the outcome may still be partly tainted by that early mistake.

A much better approach, especially with non-trivial tasks, is to behave like a golfer aiming for a hole in three. First shot: get close. Second shot: get closer. Third shot: put it in the hole.

In the context of AI, that becomes: first prompt, ask it to plan how it will approach the task. Second prompt, ask it to execute. Third prompt, ask it to review its output, as discussed earlier in this article.

Not only does this produce better outcomes, but it also helps you “debug” where something went wrong and where your instructions could have been clearer.

6) Only automate done work

Modern LLMs come with many automation tools: the so-called “Skills,” “Sub-Agents,” and more. They are wonderful. And yet, 90% of the time I have found myself building one, I would have been better served by using a simple prompt.

Don’t get me wrong. Skills, Agents, and many other advanced features are useful and can save you a lot of time. But you should only use them after you have already completed the task manually, prompt by prompt.

For example, do not create a recurring sales-report Skill before you have manually produced three sales reports with AI. First, do the work together, step by step, giving feedback at each turn. Only later, ask the LLM to turn those steps into a reusable routine.

Let me explain why this works much better. If you start by building the Skill, you will have to spend a lot of time thinking about what needs to be done and what could go wrong. Then you will have to spend time testing and debugging it. Even then, the results may not be what you imagined, and making changes will take considerable time.

Conversely, if you start by simply doing the task with the AI, you will naturally proceed step by step. As you work, you will notice what you need, what the AI needs to know, what it can do autonomously, and where it needs guidance. You will give it feedback, and it will adapt. You can continue this fast feedback cycle until you reach a satisfactory result. Only then, if you think this is a task you will need the AI to perform many times in the future, can you simply tell it: “See what we just did? Codify it into a Skill or Agent.” The AI will then quickly produce a reusable procedure that is likely to be better and more solid than anything you could have designed from the top of your head.

7) Advanced tip: manage context

Have you ever wondered why, in a long conversation, the LLM seems to lose track of where it was going, or becomes worse at doing what you want it to do?

That happens because an LLM can only hold a limited number of words in its short-term memory, which is called context. Once the conversation gets too long, parts of it get forgotten. The LLM finds it harder to focus on what matters, and performance degrades.

So, if you are working on complex tasks or expect long conversations, it is important to follow some best practices to limit context usage. The shorter the context stays, the better the LLM will perform.

• When you finish a task or subtask, start a new conversation. Before closing the old conversation, you can ask the LLM to provide a short handoff message to use at the start of the next one.
• If you send a prompt to the LLM and it replies in the wrong direction, consider editing and resending your previous prompt instead of adding a follow-up prompt. That removes the unnecessary back-and-forth from the context.
• When you need the LLM to perform a long operation where only the output matters and the intermediate steps can be forgotten, ask it to “spin up a subagent to do it.” This means that everything that happens during the task will not become part of the main conversation’s context; only the output will.
• If you use many advanced features, such as skills, plugins, connectors, MCPs, and CLAUDE.md or AGENTS.md files, know that they can considerably increase context usage, depending on how many you use and how concisely they are written. Only use what you need. For advanced users: compact your .md files and use routing tables.

---

Note: I provide workshops and office hours for teams and leaders who want to improve AI adoption without just purchasing tools that no one will use well. Contact me if interested.