I was originally going to write about something else this week.

Something about literature review with AI, probably.

But as I write this, I’m sitting at the airport in Hong Kong — midway to Nepal.

This trip wasn’t planned. My father recently fell and likely needs urgent surgery.

So here I am — traveling home, half-awake, half-worried, surrounded by the low hum of announcements in a language I don’t quite catch.

I opened my laptop, thinking I’d get some work done during the layover.

ChatGPT wouldn’t load.

“No problem,” I thought. “I’ll try Gemini.”

Didn’t work.

Then Claude.

Also nothing.

(Apparently these are banned in Hong Kong.)

So, for now, I stuck in a space where there’s no trace of AI (at least the ones I’m familiar with) — just the quiet rhythm of airport, the smell of coffee, and the realization that I’d have to write this one on my own.

It’s funny how dependent I’ve become on these tools without even realizing it.

Every time I reach for a phrase or pause mid-sentence, I can feel that reflex — the tiny impulse to ask a model to polish it.

This post, what you’re reading now, is mostly written by hand — messy, imperfect, and a little unfiltered.

But to be honest, I’m writing this rough draft with the assumption that I’ll edit it with AI later.

That, in itself, says something.

Because somewhere between these failed logins and the first paragraph, a thought started forming — one that I couldn’t shake:

The skills to write well before AI are not the same as the ones we need after it.

And like it or not, we’re not going back.

The Excavator Problem

Adam Smith, in The Wealth of Nations, said something simple but timeless: productivity isn’t about the effort you put in; it’s about the output you create.

If you can dig a foundation with a shovel, that’s work.

But if you can operate an excavator and build ten foundations in the same time, that’s progress.

No one says, “You’re not really digging.”

The skill just evolved: less brute force, more judgment, coordination, and precision.

That’s how many still talk about writing with AI.

They treat it as a shortcut instead of what it really is — a new kind of instrument.

We’ve crossed a boundary as irreversible as the industrial revolution.

You could divide creative history into B.C. (Before ChatGPT) and A.C. (After ChatGPT).

In the B.C. era, we judged writers by how well they could shovel: grammar, phrasing, endurance.

In the A.C. era, the question has shifted:

How well can you operate the machine to produce high-quality work — with clarity, ethics, and discernment?

Great writers will still stand out — but for different reasons.

They’ll be faster because they can think more clearly.

They’ll use AI as an accelerant, not a substitute.

And beginners won’t learn the way we did.

Their writing journey will begin with AI, not without it.

That’s not laziness.

That’s progress.

The Necessary Friction

But progress without friction has a cost.

Alan Cooper, a software designer, once described cognitive friction as:

“the resistance encountered by a human intellect when it engages with a complex system that changes as the problem permutes.”

Writing has always been such a system — part structure, part struggle.

That friction, the act of sitting with your own half-formed ideas, wrestling with them line by line, is what builds depth, originality, and clarity.

When AI removes all friction, it also risks removing the learning that comes from it.

The goal, then, is not to eliminate friction, but to curate it — to decide where to make things easier and where to leave the resistance in place.

Let AI handle the mechanical friction: summarizing, rephrasing, formatting.

But keep the cognitive friction that strengthens you: interpreting, prioritizing, deciding what truly matters.

Because that’s where skill is forged.

And yes, there will always be some who outsource everything to AI.

But in those cases, the output shows — the writing feels generic, thin, and devoid of real insight.

There’s already plenty of AI-generated “slop” flooding the internet — disposable text that sinks to the bottom of the barrel.

That’s exactly what will happen to unoriginal academic content as well: it may look polished, but it won’t matter.

Writing Is a Set of Microskills — and not all of it matters the same in the post-AI world

Writing well after AI isn’t just about prompting better.

It’s about mastering the 5 microskills that AI can’t replace — the ones that make your writing sharper, truer, and unmistakably yours.

These are your human edge in an AI-powered world.

1️⃣ Taste — the ability to recognize what’s good

Taste is your compass. It tells you when an idea is strong, when an argument feels elegant, or when a sentence lands just right.

AI can generate endless content. Without taste, you’ll accept mediocrity simply because it sounds fluent.

Taste comes from immersion — reading great work, dissecting why it works, and surrounding yourself with excellence.

How to train it:

Build a “swipe file” of writing you admire. Reverse-outline one piece a week. Identify what makes it flow — structure, clarity, tone, or rhythm. Over time, your taste sharpens automatically.

2️⃣ Critical Sense — knowing when AI is right, wrong, or dangerously plausible

AI doesn’t usually sound wrong — that’s what makes it dangerous.

It’s confidently plausible, even when the logic collapses underneath.

Critical sense is your internal filter — the voice that says, “Wait, does this actually make sense?”

It’s the most under-trained skill right now. And yet, it’s the one that protects you from publishing fluent nonsense.

How to train it:

Highlight one claim per paragraph and ask yourself, How do I know this is true?

If the answer isn’t immediate or verifiable, it’s not ready.

3️⃣ Synthesis — connecting fragments into coherent arguments

AI can summarize, but it can’t synthesize.

It can give you 10 points from 10 sources — but it can’t often see the pattern that ties them together.

Synthesis is what turns data into insight and text into argument. It’s where your originality shows up. It’s where you provide your unique insights about what is needed or how your work will move the field forward.

How to train it:

Before writing, outline your logic chain manually: claim → evidence → implication.

Once that’s solid, use AI to expand or clarify — not to discover your argument.

4️⃣ Judgment — deciding what to keep, what to cut, and what to question

Judgment is editing at its purest.

AI will happily produce everything you ask for — and much that you shouldn’t use.

Judgment is the discipline to say, This doesn’t serve the core message, or This sounds good, but it’s wrong.

That skill defines your credibility.

How to train it:

After every draft, ask 3 questions:

1. What’s the central point this piece makes?
2. Which paragraph weakens that point?
3. What would a skeptical reviewer challenge first? Cut accordingly.

5️⃣ Verification — ensuring what you publish actually holds up

This is the unglamorous but essential part.

AI hallucinates citations, distorts meaning, and invents details that sound legitimate.

If you skip verification, it’s not collaboration — it’s abdication.

How to train it:

Create a simple audit trail:

| Paragraph | Claim | Source | Verified (Y/N) |

Fill it as you revise. It’s tedious — but it’s how you keep trust in an age of infinite text.

These 5 microskills aren’t optional refinements.

They’re the new foundation of writing after AI.

Because AI has democratized writing speed — but not writing judgment.

Anyone can write faster now. Few can still think clearly.

That’s where your real advantage lies.

The New Kind of Writer

If you take one thing from this, let it be this:

The craft of writing has changed, but the responsibility of thinking has not.

You don’t need to compete with AI.

You need to collaborate with it — wisely, deliberately, and with judgment.

It’s not YOU replaced by AI.

It’s YOU — to the power of AI.

Those who will thrive in this new world won’t be the best prompt engineers.

They’ll be the ones who combine human taste, judgment, and verification with the machine’s speed and scale — writers who think clearly, verify relentlessly, and create meaning that lasts beyond the algorithmic churn.

Because AI can write, but it can’t understand.

It can mimic insight, but it can’t live it.

It can assemble words, but it can’t see their consequences.

And that’s where YOU remain indispensable.

Machines can mimic language. Only humans can make it matter — with their insight, lived experience, and sense of real-world consequence.

(And yes, I edited this with ChatGPT multiple times before hitting publish.)

How are you retraining your writing muscle for the post-AI world?

PROMPT OF THE WEEK

How to create scientific presentation drafts fast with AI

1. Visit Gemini and sign up
2. Upload your document (Word, PDF, or text file)
3. Upload your doc file and go to Tools > Canvas
4. Write in your prompt and press enter

Sample Prompt:

Create a concise, high-impact scientific presentation (15-20 slides) based on the uploaded document.

**Slide Deck Specs**

- Target length: **[15–20]** slides for **[12 minutes]**.
- Theme: minimalist, high-contrast, large typography, no clutter. Navy blue background.
- Accessibility: avoid red–green contrasts; define acronyms on first use.

**Required Slides (populate with content extracted from the document)**

1. **Title** — title, authors, affiliations, date.
2. **Background** — why this matters, clinical burden / gap (2–4 bullets).
3. **Objectives** — primary/secondary (label exploratory if applicable).
4. **Data & Methods** — design, cohort, inclusion/exclusion, variables, outcomes, analysis plan (tight bullets + a simple study flow graphic).
5. **Results—Primary** — key estimate(s) with effect size + 95% CI; exact p-values if provided; visual where possible.
6. **Results—Key Secondary / Sensitivity** — short table or chart.
7. **Discussion** — interpretation, what findings mean (no number restatement).
8. **Limitations** — internal/external validity; bias risks.
9. **Implications** — clinical, policy, research implications (who should do what).
10. **Next Steps** — replication, subgroups, trials, policy.
11. **Conclusion** — one-slide bullet list of key takeaways.
12. **Acknowledgments/ Funding**

**Design Rules**

- ≤6 bullets/slide, ≤10 words per bullet.
- Convert important numbers to visuals (simple tables, forest/estimate plots, flow diagram).
- Include as many visuals and figures as possible.
- **Make good use of figures from the materials provided by the user- extract and include them in the slides as they are.**
- Every figure/table has a caption and abbreviation footnote.
- Whenever possible, convert tables to easy to read figures and graphs.
- **Do NOT include figures or tables from other sources (beyond what the user provides).**
- Consistent number formatting and units.

**Rigor & Citations**

- Extract all numbers/definitions/claims from the uploaded doc.
- Mark any inferred/out-of-scope addition with an **[Inference]** tag and cite source if used.
- State analysis population (ITT/PP) and missing-data handling if present.
- Add relevant citations (Vancouver) for important claims at the bottom of the slide. No numbering.

**Guardrails**

- If an item is missing (e.g., sample size), insert **[MISSING: …]** for me to fill.
- No hallucinated stats or methods.
- Verify that all important figures from the user material are included.
- Verify that no external figures are included. 

**Deliverables**

- A completed Google Slides deck meeting the above for download.

5. Wait a few seconds while Gemini generates your presentation draft
6. Once it’s ready, click ‘Export to Slides’ to open and edit your new deck in Google Slides