You Don’t Need 100 AI Tools. You Need These 5 Skills.

The people getting extraordinary results from AI aren’t using more tools than you. They’re using the same tools differently — and the difference comes down to five things that have nothing to do with which app you opened.

I counted the AI tools I’d signed up for over the previous twelve months.

Forty-three. Forty-three accounts, most of them free tiers I’d used twice before moving on to whatever the next one was. I had a Notion page called “AI tools to try” that was longer than my actual to-do list. Every week there was a new one — a new model, a new interface, a new productivity application that promised to change how I work.

And the honest assessment of what all of that produced: almost nothing that compounded. A few hours saved here and there. Mostly, it produced the feeling of staying current without the substance of actually improving.

Then I watched someone in my field build a content and analysis system from three tools — Claude, a simple SQL environment, and Google Sheets — that was producing output I couldn’t match with my entire forty-three-tool stack. I asked them how. Their answer was not a tool recommendation. It was five things they’d gotten genuinely good at that made every tool they touched more powerful.

Photo by Brooke Cagle on Unsplash

That conversation changed how I think about the entire AI moment we’re in.

I. The tool trap — why more doesn’t compound

There’s a concept in economics called the “productivity paradox” — first named by economist Robert Solow in 1987 when he observed that the computer revolution was everywhere except in the productivity statistics. Companies were buying computers, installing software, training employees, and seeing almost no measurable improvement in output per worker. The explanation that emerged over the following decade was simple: the technology wasn’t the constraint. The capability to use it effectively was.

The same paradox is playing out in real time with AI, at the individual level.

A 2024 Nielsen Norman Group study of knowledge workers using AI tools found that the efficiency gains from AI varied by a factor of more than 10x across participants — meaning the highest performers were getting more than ten times the productivity benefit from the same tools as the lowest performers. Same ChatGPT. Same Claude. Same Notion AI. Completely different outputs.

The variable wasn’t the tool. It was the skill of the person using it.

Research from MIT’s Computer Science and AI Laboratory, published in 2023, found that the workers who benefited most from AI assistance were those with the strongest underlying domain expertise — the tool amplified what they already knew. Workers with shallow expertise got shallow results, faster. Workers with deep expertise got deep results, dramatically faster.

More tools don’t solve this. Better skills do.

The specific skills are not generic “learn to prompt better” advice. They are five capabilities that appear consistently in the people getting outsized results from AI — regardless of which specific tools they use or what field they work in.

II. Skill 1 — Asking questions that have decision value

This is the foundational skill, and the one most people skip because it doesn’t feel like a skill. It feels like common sense. It isn’t practiced like common sense.

Most people interact with AI the same way they interact with a search engine — they type a query and wait for an answer. The query is usually a description of what they want. “Write me a summary of this report.” “Give me five ideas for X.” “Explain how Y works.”

This is not a bad approach. It produces reasonable outputs. It is also the lowest-leverage way to use an AI system, because it treats the AI as an answer machine rather than a thinking partner.

The higher-leverage approach is to bring a decision rather than a query. Not “give me five ideas” but “I’m trying to decide between A and B — here’s the context, here’s what each costs, here’s what I’m optimising for. What am I missing?” Not “explain how Y works” but “I currently believe Y works this way — here’s my understanding. Tell me where I’m wrong.”

The difference is that the second approach generates outputs that are actually decision-relevant rather than informational. And it forces a quality of thinking before the interaction that produces better inputs, which produce better outputs — a compounding loop rather than a flat one.

Research by Ethan Mollick at the Wharton School found that the prompting behaviour most strongly correlated with high-quality AI output was specificity of goal, not complexity of instruction. The people getting the best results weren’t writing elaborate multi-paragraph prompts. They were being precise about what decision or outcome the output needed to serve.

Practice this week: before your next AI interaction, write one sentence finishing this prompt — “I need this output to help me decide ___.” If you can’t finish it, you’re not ready to prompt yet.

III. Skill 2 — Reading outputs for what’s wrong, not what’s right

This one sounds counterintuitive because AI outputs are often impressively good. The writing is coherent. The analysis seems thorough. The code runs.

But here’s what the research shows: AI systems are optimised to produce outputs that read as correct, not outputs that are correct. This is a different thing. An output that is fluent, well-structured, and confidently stated can be subtly or substantially wrong in ways that are difficult to catch without domain knowledge and a specifically critical reading posture.

A 2023 Stanford study examining AI-generated legal research found that large language models produced plausible-sounding citations that didn’t exist in roughly 30% of cases — and that lawyers with lower domain expertise caught far fewer of these errors than those with stronger backgrounds. The outputs looked correct. They were reading for right rather than wrong.

The same pattern appears in data analysis contexts. An AI that summarises a dataset can produce a fluent, well-organised summary that describes the data accurately in its individual claims while missing the structural issue that makes the entire analysis misleading — because the AI doesn’t know what the data is for, only what the data says.

The skill is developing a specific reading posture: trust the structure, interrogate the specifics. AI is excellent at organising information into coherent shapes. It is unreliable about specific facts, specific numbers, and specific claims in domains where it has seen conflicting training data. Read every output looking specifically for the claim that seems most authoritative and verify that one first — it’s frequently the one most worth checking.

For analysts specifically: never ship a number from an AI-assisted analysis without tracing it back to the source data manually. The AI will produce a number. The number will look correct. Verify it anyway.

IV. Skill 3 — Building workflows, not just using tools

This is the skill that separates people who save time with AI from people who save time and compound it.

A workflow is a repeatable process where each step’s output becomes the next step’s input, and the whole sequence produces a specific result reliably every time you run it. A tool interaction is a one-off transaction — you ask, you receive, you move on.

The person who has built a research-to-article workflow — a specific sequence of prompts, sources, and synthesis steps that reliably produces a publishable piece from a raw topic in a set amount of time — is not doing more AI interactions than the person who hasn’t. They’re doing the same number of interactions, but the interactions are structured so that each one feeds into the next, the quality is consistent across runs, and the time per piece decreases as the workflow matures.

Think about what this looks like for a data analyst. A one-off AI interaction might produce a useful analysis. A workflow produces that analysis consistently, at lower cognitive cost, with documented decision points that can be improved over time. The workflow is the asset. The individual interaction is not.

Building a workflow means documenting what you did when something worked well — what you prompted, in what order, with what context, producing what output — so you can reproduce it. Most people don’t do this. They have a good session with an AI tool, feel the efficiency, and then try to recreate it from memory the next time and find it doesn’t quite work the same way. The documentation is the step that converts a lucky interaction into a reliable process.

Start with one workflow. The one piece of work you do most often — the report you write weekly, the analysis you run monthly, the thing you know you’ll need to produce again. Document the AI-assisted version of it in enough detail that someone else could follow the steps. Then run it again and improve one thing. Then run it again.

V. Skill 4 — Knowing what not to delegate

This is the most counterintuitive of the five, particularly in a moment when the dominant message is about how much AI can do.

There is a category of work where AI assistance makes the output worse, not better — and it’s not the category most people assume. The failure mode isn’t usually that AI produces technically wrong content. It’s that AI produces generically correct content in situations where distinctiveness is the entire value.

Your specific perspective on a problem — built from your particular combination of experience, domain knowledge, and the specific context of your organisation — is not something AI can replicate. It can produce a competent generic version of your perspective. It cannot produce your actual perspective, because it doesn’t have access to the specific tacit knowledge that makes your perspective valuable.

This matters most in two contexts: situations where your unique judgment is what you’re being paid for, and situations where being distinctive is more valuable than being thorough.

If you’re writing analysis for a stakeholder who chose to work with you specifically — who values your read on the situation, not a coherent summary of the available data — delegating the core reasoning to AI produces an output that is professionally adequate and strategically inferior. The stakeholder doesn’t know the reasoning came from a model. But the reasoning doesn’t carry the specific context and judgment they were implicitly paying for.

The skill is being honest with yourself about which parts of your work fall into this category, and protecting those parts from delegation — not for abstract reasons about authenticity, but because delegating them produces worse outcomes, measurably.

Everything else is a candidate for AI assistance. The distinctive parts are not.

VI. Skill 5 — Learning through application, not consumption

The final skill is the meta-skill — the one that determines how fast you develop the other four.

Most people learn about AI by reading about it, watching videos about it, following accounts that cover it. This is not useless. But it is the lowest-leverage learning mode available, and the gap between what someone knows about AI and what they can do with AI is largest in people who have learned primarily through consumption.

The reason is that AI capability is largely tacit — it lives in the doing, not in the knowing. You cannot learn to ask better questions by reading about how to ask better questions. You learn it by asking questions, seeing what comes back, noticing what worked and what didn’t, and adjusting. The same is true for reading outputs critically, building workflows, and identifying what not to delegate. These are experiential skills. They compound through practice, not through consumption.

Research on skill acquisition consistently finds that deliberate practice — applied effort at the edge of current ability, with feedback — produces capability development at several times the rate of passive learning. Anders Ericsson’s work at Florida State University, spanning four decades, found this to be one of the most replicated results in the psychology of expertise. The domain doesn’t change the principle. Watching someone else use AI tools well produces almost none of the capability gain that trying to use them well yourself produces.

The practical implication is specific: every week, pick one thing you want to get better at with AI. Not a tool to try — a capability to develop. Spend thirty minutes actually doing the thing, not reading about it. Document what happened. Try to improve one specific element the next time.

Thirty minutes per week of deliberate practice compounds faster than three hours per week of consumption. The research on this is not ambiguous.

The honest close

The AI tool list will keep growing. New models will keep dropping. Every week something will be announced that seems like it changes everything.

None of that is the constraint.

The people building compounding advantage with AI right now are not doing it by having more tools or better tools. They’re doing it by being genuinely good at five things that make any tool more powerful: asking questions that have decision value, reading outputs for what’s wrong, building workflows rather than doing one-off interactions, knowing what not to delegate, and learning through application rather than consumption.

Those five skills work on the AI tools available today. They’ll work on whatever comes out next year. They worked, in slightly different forms, before AI tools existed — because they’re fundamentally skills of thinking, not skills of tooling.

Here’s where to start this week — one thing from each skill:

Skill 1: Before your next AI interaction, write the decision it needs to serve. One sentence. Do that before you open the tool.

Skill 2: Take the last thing you got from an AI tool and find one specific claim to verify manually. Just one. Build the habit.

Skill 3: Document the AI-assisted workflow you used this week. Enough detail to repeat it. That’s your first workflow asset.

Skill 4: Name one piece of work in your current role where your specific judgment — not competent generic reasoning — is the entire value. Protect that from delegation this week.

Skill 5: Pick one AI capability you want to improve. Spend thirty minutes doing it, not reading about it.

The tools are table stakes. The skills are the moat.

I write about data, AI, and careers — in plain language, no jargon. Follow for more at @data_mind.