Developers Are Writing by Hand Again

After two years of letting AI generate the code, a wave of developers is putting the tools down and writing by hand again. Not because AI failed at typing, but because something quieter went wrong: the thinking got outsourced too. Their fix is older than any model. Do the work yourself; keep AI as an assistant, not an author.

This is not an anti-AI argument. The developers leading the return still use the tools, and concede their speed. The shift is about who does the thinking. When you let a model produce something you never read, you trade comprehension for output, and comprehension is the asset that compounds. The same trade is now arriving for everyone who lets AI write their notes, drafts, and decisions.

What people believe: AI handles the typing, you keep the thinking

The comfortable story is that AI takes the mechanical part, the boilerplate and the syntax and the first draft, and frees you for the higher-order work. You stay the architect; the model is the bricklayer. It is an appealing division of labor, and for a while it feels exactly true. Output goes up. Friction goes down.

The term for the new mode is vibe coding. Rachel Thomas of fast.ai defines it precisely: "Vibe coding is the creation of large quantities of highly complex AI-generated code, often with the intention that the code will not be read by humans."¹ Read that last clause again. The dividing line is not how the code gets written. It is whether a human ever reads it — whether anyone still holds the model in their head.

The pivot: developers who tried it are walking it back

The people most fluent in these tools are the ones reversing course, and their complaint is never that AI is too slow. It is that speed hid a cost. After two years of full-time vibe coding, they found the by-hand version was the better trade once everything, not just tokens per hour, was priced in.

Developer Mo Bitar put it directly. He was "back to writing by hand for most things," he wrote, and had become "faster, more accurate, more creative, more productive, and more efficient than AI, when you price everything in, and not just code tokens per hour."²

The reversal is not one person's mood. Across the first half of 2026, the same argument crested again and again on Hacker News. Threads titled "After two years of vibecoding, I'm back to writing by hand," "I'm going back to writing code by hand," and "LLMs are eroding my software engineering career and I don't know what to do" each drew from many hundreds to over a thousand points, a sustained signal rather than a single bad week.³ The developer behind k10s.dev described the turn plainly: "This is the moment I stopped vibe-coding and started thinking."⁴

Bram Cohen, the creator of BitTorrent, named the failure mode without sparing himself the audience: "Dogfooding is when you use your own product. It's a good idea. But it can turn into a cult activity where it goes beyond any reasonable limits. In this case, the idea is vibe coding, where you make a point of literally making no contribution to what's going on under the hood, not even looking at it."⁵ The trap is not the tool. It is the decision to stop looking.

The mechanism: you lose the skill you stop using

The reason the return matters beyond code is mechanical. When you hand a cognitive task to an external aid, you stop exercising the underlying skill. Researchers call this cognitive offloading, and the evidence, though early, points one way: the offloaded skill quietly weakens. The developers rediscovered by hand what the research describes in aggregate.

In a 2025 study in Societies, Michael Gerlich surveyed 666 participants and interviewed 50 of them about AI use and reasoning.⁶⁷ The pattern was generational: "Younger participants (aged 17–25) showed higher AI tool usage and greater cognitive offloading, which coincided with lower critical thinking scores. In contrast, older participants (aged 46 and above) demonstrated stronger critical thinking skills and were less reliant on AI tools."⁸

The honest caveat matters here, and the author states it himself. This is a correlational, self-report study, a survey plus interviews rather than a controlled experiment, and it has drawn methodological criticism over its self-rated measures. It shows an association between heavy AI use and offloading, not proof that AI makes you worse at thinking. Gerlich's own framing keeps the balance: AI tools, he notes, "can enhance productivity and information accessibility," while "their overuse may lead to unintended cognitive consequences."⁹ Support, with limits, not a verdict.

The same trap is coming for your notes

What the developers learned about code is not really about code. It is about authorship. The skill you offload is the skill that fades, and thinking is a skill. The moment you let a model write the note, the summary, or the plan, you do to your reasoning what vibe coding did to the codebase: produce an artifact nobody understood.

A controlled signal from a different lab points the same way. In "Your Brain on ChatGPT," researchers at the MIT Media Lab had participants write essays with and without an LLM and found: "Self-reported ownership of essays was the lowest in the LLM group and the highest in the Brain-only group. LLM users also struggled to accurately quote their own work."¹⁰ The study is a small preprint, 54 participants and not yet peer-reviewed at capture, so read it as a clue, not a conclusion. But the clue is sharp: when the model writes it, you do not own it, and you cannot even recall it.

Notice what the developers did instead. Before any code, the k10s author wrote, "I'm doing the design work myself, by hand, before any code gets written."¹¹ Mo Bitar's first move was to "open Obsidian and begin drafting beefy spec docs that describe the feature in your head with impressive detail."¹² The pattern is identical: the return to thinking ran through a note. The plain-text file is where the human did the part the model could not be trusted to do.

There is a simple test for whether you have offloaded the thinking or kept it. Try to explain the decision, in your own words, without reopening the chat. If you can, the understanding is yours; if you cannot, you have a draft you do not own, and the MIT finding has already named the symptom. The note you write by hand is the version you can still account for next year, and accountability is the whole point.

The practice: write to think, in a note you own

The durable move is not to refuse AI. It is to keep authorship of the thinking and let the model assist the rest. The developers walking back from vibe coding did not delete their tools; they changed who holds the model of the problem. You can do the same with three habits, none of which gives up the assistant.

• Draft the thinking by hand first. Before you prompt, write the problem, the constraints, and the decision in your own words. The spec, the plan, the argument: that is the part that builds understanding, and the part worth keeping.
• Make AI assist, not author. Let the model expand, check, rephrase, or generate the boilerplate, then read every line and make it yours. If you could not quote it back, you do not yet own it.
• Keep it in plain text you control. Thinking that compounds has to be findable next year, not locked inside a chat that scrolls away. A plain Markdown note on your own device, readable and greppable and portable with no account, is where the work survives the tool that helped make it.

Frequently Asked Questions

Is AI eroding my thinking? The early evidence suggests a risk, not a certainty. Gerlich's 2025 Societies study found a correlation between heavy AI use, cognitive offloading, and lower critical-thinking scores — strongest in younger users.⁶⁸ It is correlational and self-report, so it shows association, not proof. The safer reading: the skill you stop using is the skill that fades.

Should I still do the work — write, take notes — by hand? For the thinking that compounds, yes. The developers reversing course are not anti-tool; they are protecting the part of the work that builds understanding. Use AI to assist — expand, check, generate boilerplate — but keep authorship of the reasoning. Write the plan, the argument, the decision in your own words first.

Does using AI make you worse at thinking? No study yet proves causation. Gerlich finds an association between frequent AI use and weaker critical thinking, mediated by cognitive offloading.⁹ The mechanism is plausible — unused skills weaken — but the research is early and self-reported. Treat it as a reason to stay engaged, not a reason to quit the tools.

What is cognitive offloading? Cognitive offloading is outsourcing mental effort to an external aid — a calculator, a search engine, an AI assistant — instead of doing the work in your head. It is the mediating variable in Gerlich's findings: the more a task is offloaded, the less the underlying skill gets exercised, and the weaker it tends to become.⁶

What is vibe coding? Vibe coding, as Rachel Thomas of fast.ai defines it, is "the creation of large quantities of highly complex AI-generated code, often with the intention that the code will not be read by humans."¹ The defining trait is not the AI — it is that no human ever reads or understands the output.

Why are developers going back to writing code by hand? Not for speed. They report that letting AI author code they never read cost them understanding, ownership, and the ability to maintain it — Mo Bitar found he was "faster, more accurate, more creative" by hand once everything was priced in.² The return is really a return to thinking, often routed through a hand-written plan or note.

Apps come and go, and so will every model; the thinking you did by hand is what stays, and it stays best in a note you actually own. MNMNOTE (mnmnote.com) lives in your browser as plain Markdown on your own device, no account — a place to write to think, not a tool that thinks for you.

The corpus this argument sits beside: the broader case for plain text in I Tried Every To-Do App and Came Back to One Text File, and the cognitive science of pen versus keyboard in Handwriting vs Typing: What the Research Says.