Why AI Skills Compound While One-Day Sprints Don't

# Why AI Skills Compound and a Weekend of Prompting Goes Nowhere

You blocked out a Saturday for AI. You worked through a prompt guide, tested formulas, and saved the forty best into a file. By evening, you felt ahead of everyone at work, and for that day, you were.

Three weeks later, you were working mostly the way you did before. The file is still there. You haven't opened it.

The Saturday wasn't wasted. You learned real things, and some stuck. It didn't produce the thing you were after, though, capability that keeps growing after the effort stops. There's a mental model that explains the gap. It's called compounding, and it's older than money itself.

This is article two in our mental models series. The first covered second-order thinking, the habit of asking "and then what happens?" This one asks which of your hours add up, and which ones only add.

## What does compounding even mean?

Mesopotamian scribes were solving math problems about this roughly four thousand years ago. The Babylonians had a name for it, şibāt şibtim, Akkadian for "interest on interest." The idea spent most of history as a finance concept, then traveled into general thinking through investors like Warren Buffett and Charlie Munger. Shane Parrish's Great Mental Models series later codified it as a lens for anything that grows.

The mechanism is simple. When you reinvest the gains from one round of effort into your base, the next round of gains grows from a bigger base. Growth starts feeding on prior growth.

Three things must hold for that to happen. A base that persists, so your gains have somewhere to accumulate. Reinvestment, so the gains feed the base instead of getting spent and forgotten. And time without long interruptions.

One caution. Treat this as a lens rather than an equation. Skills pay no posted interest rate and nobody audits your balance. The model borrows its language from money because money made the pattern easy to see. The transferable part is the shape, and it has one feature that matters most.

The curve is flat at the beginning and bends late. The early stretch pays almost nothing visible. A burst of effort lives inside the flat part, feels productive for a day, and leaves little behind that grows. Parrish describes what happens when people stop reinvesting in themselves, warning that "twenty years of living become the same year repeated 20 times."

That Saturday was a real deposit. It went into an account you never touched again.

## 
AI capability is a stack, not a trick

The weekend plan assumed that getting good at AI means collecting prompts. But prompting is one skill inside a set, and the set is what compounds.

The set looks like this. Writing instructions clear enough that a machine can follow them. Checking what comes back instead of trusting it, the way you'd skim an AI summary against the email thread it came from. 

Designing the workflow around the tool, deciding for instance whether it drafts and you fix, or you draft, and it tightens. Judging when AI helps and when it slows you down. And underneath all of it, knowing what you want before you type, which is a thinking skill wearing a typing costume.

The compounding lives in the connections between them. Every time you check an output, you learn where the model fails. Catch it inventing a date once, and you start pasting the source instead of trusting its memory. Every workflow you design teaches you judgment about what belongs in your hands. 

And each time you force yourself to state what you want precisely, that clarity carries over to the next tool you touch, because clarity belongs to you and to no particular tool.

Prompt formulas expire. Each one is tuned to a specific model, and new models replace old ones on a schedule nobody controls. The skills underneath ride along to whatever ships next. A file of forty prompts is inventory. The ability to write the forty-first from scratch is capital, and capital is what compounds.

## But beginners gain the most, don't they?

A famous research finding seems to point the other way. Economists Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied thousands of customer support agents at a large software firm and found that access to an AI assistant helped the least experienced agents the most, while the most experienced saw little benefit.

That finding is real, and it fits. It describes borrowed gains, the tool lifting your output while you use it, and borrowed capability goes home when the task ends.

A second line of research shows what happens when borrowing replaces the learning itself. In randomized experiments published in January 2026, Judy Hanwen Shen and Alex Tamkin, researchers at Anthropic, had experienced developers learn an unfamiliar programming library. 

The group working with an AI assistant scored worse afterward on understanding the library, reading its code, and debugging it, and on average they finished no faster. The participants who delegated everything to the AI finished quickest and learned the least. The work got done. The base didn't grow.

The same experiments carry a hopeful detail. Some heavy AI users still learned well. They asked conceptual questions, requested explanations alongside generated code, and stayed mentally in the loop. Engagement separated the learners from the non-learners more than AI use did.

Researchers at Harvard Business School and Stanford found a related boundary. Iavor Bojinov, Edward McFowland III, and colleagues watched workers at a trading firm use AI for writing tasks outside their specialty. AI closed the gap on organizing and outlining, and couldn't close it on execution when the person's expertise sat too far from the task. As the researchers put it, "GenAI can provide the map, but navigating the terrain is another matter."

The three findings tell one story. AI will boost your output from the first day, and that boost is worth having. It can't deposit skill into your account on your behalf. The weekend prompter made one mistake, treating a rental as a purchase.

## One task, watched for a month

Watch the mechanism run on something ordinary, a weekly status update, with whichever AI assistant you keep open.

Rep one. You paste your raw notes and ask for a short update. The draft reads fine until you notice it buried the one item your manager cares about. You just trained checking.

Rep two, a week later. You add one line before the notes, saying who reads this and what they need first. The draft comes back with the right emphasis. That line is instruction writing, learned from last week's miss.

A few weeks in, your one line has grown into a short pattern you paste every time, covering audience, priorities, length, and what to leave out. It keeps fumbling numbers copied from your notes, so you verify those by hand and stop re-checking what it reliably gets right. That split is judgment forming.

A month in, the update takes ten minutes instead of forty. The same short pattern now writes your meeting recaps too, because clarity carries between tasks. No week looked impressive. The stack was assembling across weeks, which is what the flat part of the curve feels like from the inside.

## What reinvestment looks like

The practice that feeds this loop is small enough to hide inside a normal day.

Pick one task you already do, a recap, a first draft, a rewrite, and run it through your assistant as part of doing it. Real tasks force the whole stack at once, stating what you want, checking the result, deciding what stays in your hands, in a way that practice exercises rarely do.

Then spend one minute noticing. What it got wrong, what your instruction missed, what you'd phrase differently. That minute is the interest payment going back into the principal. Skip it, and the rep still helps you in the moment, then evaporates.

Give the gains somewhere to land. A running note in whatever app you already use, holding patterns that worked and failures you've caught, turns loose lessons into a base. Next month starts from this month instead of from zero.

And borrow the habits that kept learners learning in the skill-formation research. Ask the tool to explain its answer, then read the explanation. Try your own version first on anything you want to get better at, then compare. Make it show its steps instead of a finished conclusion. Each keeps your head in the loop, which decides whether a rep feeds the base.

Twenty minutes on most days will pass any single heroic weekend. The curve holds no opinion about discipline or moral credit. Steady reps are the only input the mechanism accepts.

One reassurance, because the early stretch is where most people quit. If the gains feel small a few weeks in, nothing is wrong with you. The front of the curve is flat for everyone. The people who look effortless with AI stood where you are, one unremarkable rep at a time, until the bend arrived.

## The model in one breath

Capability with AI grows the way anything compounds, on a base you keep feeding. A burst buys you a day of borrowed strength. Steady reps, attached to real work, with your head in the loop, buy you a curve. Then time does the part that no single weekend can.

More mental models are coming in this series. The previous article taught you to ask "and then what?" This one is the answer. It adds up.

## Sources

- Shane Parrish, The Great Mental Models, Vol. 3 (Systems and Mathematics), Farnam Street: https://fs.blog/tag/compounding/

- Kazuo Muroi, "The Oldest Example of Compound Interest in Sumer" (şibāt şibtim, Old Babylonian mathematics): https://arxiv.org/pdf/1510.00330

- Erik Brynjolfsson, Danielle Li, Lindsey Raymond, "Generative AI at Work," Quarterly Journal of Economics: https://academic.oup.com/qje/article/140/2/889/7990658

- Judy Hanwen Shen, Alex Tamkin (Anthropic), "How AI Impacts Skill Formation" (January 2026): https://arxiv.org/pdf/2601.20245

- Ben Rand, "Gen AI Boosts Productivity, But Can't Turn Novices Into Experts," HBS Working Knowledge, on the working paper "The GenAI Wall Effect" by Vendraminelli, DosSantos DiSorbo, Hildebrandt, McFowland, Karunakaran, and Bojinov: https://www.library.hbs.edu/working-knowledge/gen-ai-boosts-productivity-but-cant-turn-novices-into-experts

If you found this useful, check out my newsletter below

I share one AI superpower every week

Subscribe, it's free

https://linktr.ee/alex_prompter