A blog post titled "AI is a tool and using it well takes skill," published by Yyhh, climbed the Hacker News front page this week, with 287 points and 238 comments at last count. It pushes back on Rich Hickey's claim that LLMs are automation rather than tools. The author's case, as summarized: AI can be wielded skillfully when a human supplies judgment, feedback, and oversight.
Whether or not you land on the author's side of that argument, the practical question is the useful one. If the output depends on the person using it, what should that person actually do differently? Here's a working answer you can apply to your next session with a chatbot or coding assistant.
Why the tool-versus-automation distinction matters
Per the post, the author frames LLMs as instruments that amplify the user's habits, standards, and expertise rather than replacing them outright. That's a claim worth testing against your own experience.
If an LLM is automation, you hand it a job and accept the result. If it's a tool, you remain the one doing the work, and the quality of the result tracks the quality of your direction and your review. The difference changes where responsibility sits. With a tool, accountability stays with you.
That's my read of the implication, not a quote from the post, but it's a useful test: if you can't explain and defend what you shipped, you used it like automation.
Treat the output as a mirror of your standards
One way to read the author's argument is that AI reflects the person at the keyboard. Sloppy habits in, sloppy results out. Careful standards in, noticeably better results out.
Try this:
- Write down your standard before you prompt. What would a good answer look like? What would make you reject it?
- Give context you'd give a new colleague. Goals, constraints, audience, what you've already tried.
- Check the result against your criteria, not your first impression. Fluent text feels right even when it isn't.
The skill here is mostly the old skill of knowing what good looks like. AI doesn't supply that for you.
Use it for learning, critique, and iteration
The strongest uses, in the spirit of the post's emphasis on feedback and oversight, are the ones where you stay in the loop:
Ask it to critique, not create
Write your draft, your function, or your research summary first. Then ask the model to find weak spots, missing cases, or unclear passages. You decide which critiques hold up.
Ask it to explain, then verify
When you're learning something new, have it explain a concept in plain terms, then ask for the strongest objection to that explanation. Check anything important against a primary source.
Iterate in small steps
Instead of requesting a finished product, work in short rounds: outline, review, revise, review again. Each round is a chance to apply your judgment, which is the part the author argues can't be skipped.
A workflow for programmers and researchers
For programmers, a habit that keeps you in charge:
- Describe the problem and your intended approach yourself.
- Ask for options or a review of your approach, not a full solution.
- Read every line of any code you accept. If you can't explain it, don't merge it.
- Test it the way you'd test a teammate's code.
For researchers, the same shape works. Form your question and a tentative view first. Use the model to surface counterarguments or related angles. Then confirm claims and citations yourself before they go anywhere near your work.
Neither workflow is exotic. Both put your judgment at the start and the end of the process, with the model in the middle.
Don't treat it like a person
Anthropomorphizing is one of the angles floating around this debate, and it's a practical hazard as much as a philosophical one. When a chatbot sounds confident and friendly, it's easy to extend it the trust you'd give a knowledgeable colleague. A habit that helps: describe what it did in tool terms. It produced a draft. It generated a suggestion. That wording keeps the question "is this right?" in front of you.
Why this debate is landing now
AI is showing up in writing, coding, research, and education. The conversation is moving from "should we use it?" toward "what counts as responsible use?" The post stakes out a middle position, pushing back on both hype and outright rejection by focusing on skill, agency, and accountability. The 238-comment thread suggests plenty of readers have strong opinions about where that line sits, and I'd expect reasonable people to draw it differently.
Your next move
Pick one task you currently hand to an AI wholesale, and rebuild it so you write the standard first and review the result last. If the output gets better, you've just seen the post's argument play out on your own screen.
