Be Productive with AI: Context. Goal. Access. Let It Do the Work.

Last updated: 15.09.2026 08:00

AI is a great tool to support your work. If you want to boost your productivity, your first thought is usually to improve your prompts. But small, key changes can have a much greater impact. Self-analysis can help.

To find out how I work with AI and why I often get more done than others, I analyzed over 800 of my messages from several months of platform development using an AI solution. The result:

• 64 percent of the messages are shorter than 200 characters; the average is 120 characters.

• 37 percent contain a question mark—so I ask a lot of questions.

• 19 percent essentially just say “continue.”

• Just 11 percent match what one would typically think of as a prompt: longer tasks with context and requirements.

If just 11 percent of the messages correspond to a classic prompt, that means that in nine out of ten messages, no new, clear task is given. Instead, I ask for an update, give approval, set a limit, question a result, or simply forward the message. A pattern emerges:

1. I set the context.

2. I set a goal

3. I provide access rather than instructions

4. And then I step aside.

1. Setting the Context

An LLM can figure out almost anything on its own—except for things known only within your own company. That’s exactly where—and only there—humans are irreplaceable. Language models are good at language, but they know nothing about your business. LLMs don’t know that a certain customer group is at risk of churning. That a certain phrasing will set off the legal department. That a certain customer tends to be sensitive. That’s exactly where your contribution lies, and it’s worth far more than any nuance in the prompt wording.

In practice, this means:

• Provide the context that only you know. And the shortest corrections are the most valuable. My most effective messages aren’t rephrasings. They’re facts I add later, for example:

“Wrong. The component is also on the customer side; that’s how customers expand our platform.”
“Semantic checks must be multilingual; otherwise, there’s no point in doing them at all.”
“The interfaces don’t matter; there’s no user on the integrations.”

None of these messages criticizes a result. On the contrary: each one provides additional information that the AI model couldn’t have known. And as soon as it knew that, the rest fell into place on its own. In practice, this means:

• Turn every correction into a fixed rule. That means: If I have to correct the same mistake twice, the first correction was worthless. That’s frustrating. Insight: A correction that I merely voice is a conversation. A correction that I write down is a check. My corrections have now turned into over 142 written rules that the system reads aloud at the beginning of each session. One of them came about like this:

“No half-measures, please. This applies not only here, but in general.”

Since then, this sentence has applied to every delivery, not just the one that prompted it. In the contact center, this means: Your prompt library, your guidelines, and your quality management criteria aren’t just paperwork—they’re the memory of your operation. Everything that exists only in the mind of an experienced team leader will be lost with the next model change, the next vendor change, and the next job change!

2. Set a Goal

Write down the goal, not the next step! My most frequent message by far is actually my most useless one. 155 times (19 percent of messages), I essentially just wrote “continue.” So 155 times, the work came to a standstill and the AI waited for a decision. And decisions get put off. I ended up making them with a single word. But if one word is enough, then that decision never needed me. It needed a document. The solution was right there in the same paragraph: a prioritized list in a file, so that continuing the work means reading the top line instead of rethinking things.

In practice, this means:

• At the start of a longer project, I have the model write the plan itself into a file. I read it, correct the premises (see Habit 1), and approve it. And then the task is no longer “take the next step,” but rather: “this document has been fully implemented.”

“Write the plan into a file. The goal is to have it fully completed by tomorrow morning.”
“What’s still left to do in the plan?”

This way, my own workload shrinks to four steps: The model writes the plan. I correct the premises and approve it. A command sets the goal. I walk away. So now, for the 155 “next steps,” there’s simply a command. This document must have three characteristics; otherwise, it won’t work:

• The document belongs to the model, not to you. Let the model write the plan and confirm it, rather than dictating it. A plan that you formulate makes you the bottleneck again—just one level earlier.

• It has a verifiable “completed” state. This involves three things: a measurable end state, specified evidence of it, and the constraints that apply along the way. If these are missing, it’s not a target—it’s a wish list.

• The document is updated regularly. The current status belongs in the document, not in the conversation log. Only then will the assignment survive a system crash, a model change, and the night.

This applies to your business just as much, with or without AI. “Work through the queue” is a “keep going”: It ties you to every repetition. A documented target state runs without you, something like this: these forty transaction types will be classified by Friday, the disputed cases will be identified, and the status will be recorded in the document.

3. Grant access instead of giving instructions

Open the door! Every time an AI isn’t allowed to check something on its own, a task turns into a query. And every query gets left unresolved. So whoever denies access isn’t in control—they’re making themselves the bottleneck. A surprising number of my messages open a door:

“The access credentials are in the Env file. Check the logs yourself.”
“You can log in there yourself; check it directly in the system.”
“With 570 connectors, I can’t be the bottleneck. Find a way to do it without me.”

In the contact center, therefore, the crucial question before any AI project is: What is the system allowed to access on its own? An AI that isn’t allowed to read the knowledge base, the case history, and the rate data on its own may, at best, phrase things more elegantly, but it doesn’t relieve anyone’s workload. That’s because, as before, a human still has to compile every single piece of information.

In practice, this means:

• Define the boundaries, not the path. Instead of telling people exactly how to do something, specify what’s allowed to happen in the worst-case scenario. Autonomy is affordable as long as the damage is limited. 7 percent of my messages are therefore purely about setting boundaries:

“Everything in the data service, but don’t touch the shared library.”
“Nothing in production yet. Dev only.”
“For registrations, use only the shared mailbox—never a private address.”

Let the AI classify, draft responses, summarize conversations, and research information. And define the handful of actions it must never perform on its own: credit notes, cancellation confirmations, and legally binding commitments. Freedom within clearly defined boundaries yields more than constant oversight everywhere.

4. Get Out of the Way

This is the hardest part, but the one with the greatest impact. If you stay out of it, it’s easy to lose touch with whether things are actually moving forward. But “stepping aside” doesn’t mean looking the other way. That’s why reviewing the results is the only task I won’t let anyone else handle.

In practice, this means:

• “Done” is a claim, not a state of being. Thirteen percent of my messages therefore question a reported result:

“Is that in the interface? Has the plan really been completed? Has it been fully tested?”
“Did we actually demonstrate this, or are we just assuming it?”
“Has it been rolled out yet? Nothing’s happening on my end.”

So never ask for the report. Ask for the result. Has a real process run its full course? From the incoming customer message to the response that reaches the customer. Not in testing, but in production.

Insight & Conclusion: More Successful Without Me

Greater productivity doesn’t come from a better prompt. It comes from three actions: setting the context, providing the goal and access, and getting out of the way. The first two take time at the beginning. The last one costs you the feeling of being needed. But it’s worth it. For months, my most frequent message was “continue.” Every single one of them highlights a point where I was getting in the way myself. When you step back, you get more in return than when you interfere.

Tip: Try measuring yourself!

Read through your last fifty messages to an AI system and sort each one into one of six categories:

• new task,

• premise, rule,

• goal,

• approach,

• Continue.

If almost everything falls under “new task,” then you’re still using a tool instead of leading. If “carry on” is the largest category, then you’re missing a written goal. These two numbers say more about your future output than any prompt guidebook.