Five ways to plan a week with AI

This piece first appeared in German. The English version is a rewrite rather than a line-by-line translation, and the German original stays online in the archive: Zeit- & Projektmanagement mit KI.

A working student in consulting opens her laptop to twelve emails, a thirty-page client briefing, a presentation due Friday and a pile of internal admin. Work down that list by urgency and the day fills up without any of the strategic work getting done.

The five methods below change what the model is asked to process. Each one makes it weigh something beyond the text: stakeholder interests, logical structure, risk, contradiction. They apply to university work and professional work alike.

Priority is a multi-dimensional question

Tasks compete for attention, and sorting by deadline alone produces a lot of completed items and no progress on anything that matters.

Extend the prompt into a multi-dimensional assessment, and the model becomes a strategic filter that weighs tasks by impact, energy cost and stakeholder value, then places them against your own concentration curve.

Here is my task list: [list]. Build a prioritization matrix. Score each task from 1 to 10 on strategic impact for the client, impact for the team and my manager, energy required, and deadline risk. Account for the fact that I concentrate best in the morning: schedule high-energy tasks before 12:00. Then produce a timetable for today, with the reasoning.

What comes back looks like this:

The reasoning matters as much as the grid. Briefing and presentation carry the highest cognitive load and the largest strategic impact for both stakeholders, so they belong in the window where concentration peaks. Admin has almost no client impact and scores 9 for the manager, which makes it a low-energy task for the afternoon dip.

Three things become visible. Deep work belongs in the morning, because that is the most valuable resource on the list. Neither the client nor the manager gets neglected. And the model acts as a corrective against spending the best hours on easy, low-impact email.

Plan backwards from the deadline

Large projects feel diffuse and distant, which is where the planning fallacy lives: the time actually required gets underestimated and the start gets postponed.

Force the model to think from the hard deadline backwards, never from the start date, and the real time pressure becomes mathematically visible.

Today is 1 March and the submission is due on 22 March. Plan my project backwards from the deadline. Define the latest possible day for a complete rough draft and the deadline for finishing the literature search. Build in a buffer of 20 percent for illness or writer's block, and give me weekly milestones.

For a forty-page seminar paper in twenty-one days, the answer lands roughly here. Final proofreading and formatting occupy the last three days, inside the buffer zone. The rough draft carries a hard deadline of 15 March, and the literature search and outline finish a week before that. Week one goes to structure and sources, week two is the writing sprint, week three is polish.

The student sees immediately that three weeks is not what he has. He has about ten highly productive days. Procrastination becomes arithmetic, because the buffer sits at the end of the plan as an untouchable block.

The Feynman iteration works on any subject

The principle holds for quantum physics, tax law and learning a new piece of software equally well. Take a junior data analyst who has to understand Monte Carlo simulation.

Use the model as a sparring partner that leads you deeper in stages and tests your knowledge with deliberate traps. Phrase a question uncritically and these systems will happily confirm a false conclusion, and that is exactly why the critical framing matters.

Explain [topic] so that a ten-year-old understands it.

Now explain the mathematical and technical foundations underneath, and the principles that follow from them.

Give me a short exercise with deliberately built-in trick questions, answerable correctly only by someone who has actually understood the topic and not by someone who memorized it.

For Monte Carlo the three answers run something like this. Imagine you want to know how often a die comes up six, so you roll it ten thousand times on a computer and get a realistic distribution. Then: the method rests on random samples from a probability distribution, used to compute numerical approximations of complex integrals. And the trap: if you double the number of simulations, does the error rate halve? It does not. The error falls by a factor of the square root of two.

The learner moves through three stages: simplification to remove the barrier, theory to consolidate the knowledge, and application under test conditions.

Trick questions simulate a real examination, which moves the learner out of passive consumption and into actively defending what they know.

Assign the model the role of the critic

Writing on your own leaves out the change of perspective. The argument stays coherent inside your own logic while structural weaknesses, implicit assumptions and attackable premises go unnoticed.

Prompted properly, these systems can occupy different intellectual positions and examine an argument as though the criticism came from outside. Assign the role explicitly: skeptic, critic, methodological examiner. Leave the technical vocabulary out, so the assessment rests on the structure of the argument, on causality and on the logic of the justification, and not on authority or density of jargon.

Take the perspective of a critical outsider. Which assumptions are implicit in my thesis? From which three perspectives could this argument be attacked? How would it have to be formulated to survive those attacks?

The model becomes a simulated counterpart with shifting viewpoints, and it shows where an argument works only inside one system of thought and where it holds up from outside. Arguing well means anticipating criticism more than being right, which turns the tool into an instrument for perspective and away from a confirmation machine.

Make the trade-off explicit

Decisions get made on instinct more often than anyone admits. Take a startup choosing between an AI chatbot and a data protection dashboard as the next feature on its website.

Set the options against each other.

Compare feature A [description] and feature B [description]. Build a decision matrix covering revenue potential, technical effort and risk. Run a SWOT analysis for each. Which feature wins if our goal is fast market growth?

The analysis usually surfaces something the instinct missed: feature A offers the larger opportunity, and feature B removes a critical legal risk. Either way the decision becomes rational and, just as importantly, documented in a form that investors and managers can follow.

Friction sharpens the output

All five methods do the same underlying thing. They force the model to process metadata, meaning stakeholder interests, logical structures, risks and contradictions, where it would otherwise assemble text blocks. The more friction you create through counter-questions and scenarios, the sharper what comes back.

This piece is part of a workshop series.

Sources

Grouped by the section they support.

Opening

• No external source. The example is the author's own.

Priority is a multi-dimensional question

• No external source. The prompt and the sample matrix are the author's own.

Plan backwards from the deadline

• Buehler, Griffin and Ross (1994), Exploring the "planning fallacy": Why people underestimate their task completion times, Journal of Personality and Social Psychology 67(3), 366–381, doi:10.1037/0022-3514.67.3.366. Supports the planning fallacy.
• The sample plan is the author's own.

The Feynman iteration works on any subject

• Sharma, Tong, Korbak, Duvenaud, Askell et al. (2023), Towards Understanding Sycophancy in Language Models, arXiv:2310.13548. Supports the tendency of language models to agree with a user's false framing.
• Owen (2013), Monte Carlo theory, methods and examples, chapter 2, artowen.su.domains. Supports Monte Carlo estimation and the error falling with the square root of the number of samples.

Assign the model the role of the critic

• No external source. The method is the author's own practice.

Make the trade-off explicit

• No external source. The example is the author's own.

Friction sharpens the output

• No external source. The conclusion is the author's own.