No AI Fridays: one day a week writing code by hand

A single-page manifesto called No AI Fridays is making the rounds this week, and the pitch is exactly as blunt as the name: one day a week, "turn off your AI assistants for the day, write code with your own hands, read the documentation, and think things through yourself." The site's author writes that "as the CEO of HTMX, I've mandated No AI Fridays" at his own company. We run a studio that ships agent-built software for a living. We think the argument is worth taking seriously anyway.

What actually happened

The site is short and argues one thing: leaning on assistants full-time accrues what it calls cognitive debt. It lists reduced work engagement, weakened critical thinking, and impaired skill development, plus a blind-spot problem — you stop noticing which decisions you are no longer making. It also notes a side effect that will please your finance person: fewer tokens burned.

The carve-out is the part most people miss. Feedback tools stay on. The site names greptile and prelint explicitly, and gives the reason: "These are actually awesome when used with handcrafted code — they give you feedback that you can learn from." The distinction is generation versus critique. A linter or a review bot tells you something about code you wrote. A generator hands you code you did not.

The format is an FAQ, not a rulebook. It allows for more than one day a week, and for teams that go seven.

Why AI-free practice matters for your business

Here is the version that survives contact with a P&L, because "developers should suffer more" is not a business case.

You are one incident away from needing the skill. The agent writes the migration. Then production is down at 11pm, the assistant is confidently wrong about your schema, and someone has to read the query plan themselves. That capability does not survive on its own. Teams that never practice it discover its absence at the worst possible moment.

Review quality is the real bottleneck, and review is a skill. If your engineers are shipping more code than they can meaningfully read, the constraint on your delivery is not generation speed — it is judgment. A day spent writing by hand is a day spent recalibrating what good looks like, which is the thing that makes the other four days safe.

It is cheap to test. Pick a Friday. Run it once. Note what people reach for and cannot find, and what took four hours that should have taken forty minutes. If nothing surfaces, you have learned your team's fundamentals are solid. If a lot surfaces, you have found a risk that was already there.

We are not going to stop building with agents, and neither should you. But a practice that costs one day and exposes where your team has quietly stopped knowing things is a good trade at this price.

Sources: No AI Fridays.