The LLM Critics Are Right. I Use LLMs Anyway: A Guide

Here’s an unpopular position: the loudest LLM critics are basically correct, and you should still use these tools every day. Not despite the critiques – because of them. Once you accept that LLMs produce a lot of slop, flatter you constantly, and depend on a shaky geopolitical stack, you start using them completely differently. Better, actually.

This post is a practical follow-up to Jeremy Theocharis’s essay “The LLM Critics Are Right. I Use LLMs Anyway”, which hit the Hacker News front page around mid-November 2026 and set off a wave of “finally, someone said it” replies. His piece is a manifesto. This one is the workflow.

The one-paragraph context you need

Theocharis’s core claim is that LLMs amplify what you already have: opinions, structure, frameworks. “If you have thoughts, they come out sharper and faster. If you have nothing, nothing comes out, very fluently.” He accepts the standard critiques as true – copyrighted training data, environmental cost, ethical problems, and the NVIDIA/OpenAI money-moving circle that probably ends in a burst bubble – and still uses LLMs constantly. The trick is knowing exactly what job to give them.

The four-step amplifier workflow

Forget the “write me a blog post about X” pattern. That’s the pattern that produces slop. Try this instead – it’s the workflow I’ve converged on after reading the Theocharis post and a few adjacent ones.

1. Input: You write the first draft. Ugly, rough, half-formed. But it’s yours. Bullet points are fine. Fragments are fine.
2. Critique: Paste it into the LLM and ask for holes, contradictions, missing pieces. Do not ask it to “improve” anything yet.
3. Amplify: Take the critiques you agree with. Rewrite the draft yourself, integrating them. Ask the LLM to sharpen specific sentences you flag.
4. Strip: Delete anything the LLM added that doesn’t sound like you. If a sentence feels like it could belong to any tutorial on the internet, cut it.

The whole point is that the LLM never generates the primary artifact. It only pressures your existing artifact into a better shape. Theocharis puts it this way: he uses it to make fewer things more high quality, and tends to use an extreme amount of tokens just to prepare a couple of sentences for a human.

The critique prompt that actually works

Most people ask “what do you think of this?” and get flattery. That’s because LLMs are RLHF-trained to please you. The workaround, from Umang Jaipuria’s “21 ways to use LLMs better”: remove yourself from the prompt entirely. Present your ideas as someone else’s work and you’ll get less flattery and more candor. Here’s the version I use:

A colleague sent me the draft below and asked for brutal feedback
before they publish it. They said flattery is not useful.

List:
1. Every claim that is unsupported or vague
2. Every sentence that could be cut without losing meaning
3. Every place the argument contradicts itself
4. One thing the piece is missing

Do NOT rewrite anything. Just list.

---
[paste your draft]

Two things matter here. First, “a colleague sent me” – you are not the author in the prompt. Second, “Do NOT rewrite anything” – this stops the model from generating a shiny replacement version that you’ll be tempted to just paste in.

Why you should chunk aggressively (the lost-in-the-middle trap)

People assume that because Claude and GPT-5 have huge context windows (as of 2026), you can just dump a whole document in and ask questions. This is wrong in a specific, measurable way.

Turns out the U-shape is the real story here. The Stanford “Lost in the Middle” paper (Liu et al., arXiv:2307.03172) found that LLM accuracy degrades – falls off hard – when relevant information lands in the middle of a long context. The beginning and end of your prompt get processed reliably; the middle doesn’t. Your 200K-token context window is basically a stage: the front and back rows get lit, the middle sits in shadow.

Practical rule from the Janea Systems production playbook: chunk into 400-800 token segments and summarize at transition points between workflow stages. For the four-step workflow above, this means: don’t paste a 20-page PDF as one blob. Split it into sections, critique each, then merge.

Pro tip: When you must use a long context, put the thing you actually care about in the last ~10% of the prompt. It’s the recency slot. Everything before it is context; the tail is where the model “looks hardest.”

Common pitfalls

The traps I keep watching people fall into – including me, earlier this year.

• Trusting fluency as a signal of correctness. A confident, well-formatted answer is not a correct answer. This is the whole “looks good, nobody would read out loud” problem Theocharis names: LLMs are good at producing massive amounts of content that looks good, but that nobody would read out loud in front of an audience.
• Asking for opinions in fields with strong existing consensus. The model defaults to the most popular technique. Theocharis: “in fields with a lot of opinions, the LLM will just tell you what most people would want: the most popular technique, which is maybe not the best one for your case.”
• Locking your workflow to a single provider. On June 12, 2026, a US export-control directive forced Anthropic to abruptly disable its frontier models for all non-US customers. If your business ran on that, it broke overnight. Have a fallback.
• Trying to contribute LLM-generated PRs to open source. Don’t. Projects like Zig and Gentoo are already refusing to accept LLM-generated PRs – and even Armin Ronacher, creator of Flask and founder of Pi.dev, said live on stage at Local-First Conf Berlin that they auto-close almost all incoming LLM submissions on Pi.dev itself.

What the numbers say

There’s a gap between what LLMs can do inside a company that has committed to the tooling, and what they can do for the rest of us. The gap is smaller than the hype suggests and larger than the critics claim.

When NOT to use this

• Original research where you have no prior view. The model will hand you the majority opinion of its training data, dressed as neutral summary. You’ll mistake that for consensus.
• Anything requiring uncommon, low-probability solutions. LLMs predict the statistically likely next token. That’s the opposite of “clever.”
• Work you can’t verify. If you can’t check the output against reality – code that runs, a fact you can look up, a person who can read it – you’re gambling.
• Written communication that is supposed to be from you to another human. Theocharis is direct: written text should be from humans to humans. Use the LLM to sharpen your draft. Don’t let it ghostwrite.

FAQ

Isn’t this just fancy prompt engineering?

No. It’s a role reassignment. Prompt engineering optimizes what the model produces. This workflow constrains the model to a supporting role and keeps you as the author.

What if my draft is really bad – won’t the LLM just tell me it’s fine?

Yes, unless you strip yourself out of the prompt. If you write “here’s my draft, be honest,” you’ll get a compliment sandwich. Write “a colleague sent me this and asked for brutal feedback” and you get real critiques. Same model – the difference is entirely in whether it thinks it’s being polite to you. Try both on the same text and compare. The gap is bigger than you’d expect.

Is the bubble really going to burst?

Probably, in some form. But bubble bursts don’t delete the underlying technology – they reprice it. Even after a crash, the models will still be here, the workflows will still work, and the people who learned to use them as amplifiers instead of oracles will keep shipping.

Next action: Open a document you’ve been avoiding. Write the ugliest 200-word draft you can in ten minutes. Then run it through the critique prompt above. See what comes back.