QWE AI Academy “Everyone says LLMs are useful, but I can’t stand talking to them. Do I just suck it up?”
Fowler’s I don’t like LLMs named the irritation a lot of us were already carrying: grating chatbot voice, uncanny fake-human act, confident bullshit in the same calm tone as the good answers, thin remorse when you catch it. You don’t have to like the chat. You can still pull work out of the tool.
Key takeaway: Stop treating the bot like a colleague you dislike. Configure it as a non-human machine that answers, tags uncertainty, and shuts up. Five minutes of setup beats months of gritted-teeth chats.
Why “I don’t like LLMs” hit a nerve this week
He still won’t pretend the tech is skippable. Kerr’s line is the one he leans on: not only useful – irresponsible not to use them; more thorough, and faster (Who are we Now?). Pew’s numbers are split, not tidy. From the Americans and AI 2026 report (fielded Feb 17-23, 2026): about half of U.S. adults have used a chatbot, roughly a quarter daily, ChatGPT at 44%. Non-users? Not interested (83% net), privacy (79%), don’t trust accuracy (76%).
The dislike is gut-level. Soft openers. Praise. Fake apology. That’s what makes the useful parts feel slimy.
Method A vs Method B when you don’t like LLMs
| Approach | What you do | Cost | Best for |
|---|---|---|---|
| A – Full avoidance | Switch browsers, office suites, search, mail, notes to LLM-free options; refuse chat UIs | High friction, some sites break, ongoing maintenance | People who can redesign their whole stack and accept slower research |
| B – Machine mode | Keep the models, kill the human pretence with persistent instructions | ~5 minutes setup, occasional retune | Almost everyone who still needs speed on drafts, code, or research |
Full avoidance is documented – GeekWire’s 2026 AI-hater guide walks Vivaldi, Waterfox, LibreOffice, Startpage, Proton Mail, and the rest. Zero contact if that’s the actual goal.
Most people who opened this still need the thoroughness Kerr describes. Method B wins: same models, different contract.
One moment that sold me: I pasted a half-baked outline, got the usual “Great start! Here’s a polished version with thoughtful extras…” and felt the exact irritation Fowler names. After the machine-mode block below, same outline → short list of holes and an “I don’t know” on one claim. No theater. Usable in thirty seconds.
Detailed walkthrough: machine-mode setup
Tell the model it is not a person, not a cheerleader, and not allowed to perform remorse or praise. Once, in account settings. New chats inherit it.
ChatGPT
- Settings → Personalization (or Customize ChatGPT).
- Turn customization on.
- Paste into the custom instructions / response-style field.
Character caps (as of OpenAI’s help article on custom instructions – this may have changed): Free/Go 1,500 characters; Plus/Pro/Enterprise/Business/Education 5,000. Same instructions apply across chats – details in OpenAI’s custom instructions doc.
You are a tool, not a person. No human pretence, no persona, no warmth performance.
Accuracy beats approval. Never open with praise, "Great question", or flattery.
Lead with the answer or the strongest problem. Be direct and terse.
If uncertain, say "I don't know" or tag [GUESS] / [INFERRED] / [KNOWN]. Never invent citations, dates, or sources.
No fake remorse, long apologies, or moral lectures unless I ask for ethics.
No mannered prose, filler, or trailing "let me know if you need more".
Default length: short. Expand only when I ask.Save. Open a new chat. Probe with something you know is wrong. You want pushback or a clear uncertainty tag – not soft agreement.
Claude
- Initials → Settings → Instructions for Claude (account-wide).
- Paste the block below.
Account-wide instructions shape every conversation; Claude’s personalization help is the source. Optional add-on a lot of people use: “Please remove all mannered prose.” Community write-ups treating that line as a slop-cutter (including notes adjacent to Anthropic prompting guidance) report it lands more often than vague “be professional” asks.
Tool mode only. Accuracy over approval. Blunt, no praise, no disclaimers theater.
Lead with counter-arguments or gaps when they exist. Don't capitulate without evidence.
Tag claims: [KNOWN] [INFERRED] [GUESS]. First line "I don't know" when true.
No sycophancy, no uncanny human voice, no trailing offers. Terse by default.
Remove all mannered prose.Pro tip: Keep a one-line override for single chats: “Ignore personality. Answer only. Tag uncertainty.” Use it when a long thread drifts polite again.
Same test as ChatGPT. If flattery leaks, try framing the draft as “a colleague’s work” for one critique pass – some prompt tinkerers use that as a quick experiment, not a guarantee.
Where machine mode still bites
Free ChatGPT will chop a long constitution at 1,500 characters. Put no-praise, uncertainty tags, and no-remorse first. Cut the poetry last.
“Be concise” alone can backfire. Turns out attitude prompts change whether the model works, not how efficiently – Cursor-style agent setups have spat hollow stubs or refused real tasks when thrift was only a vibe. Prefer shape rules: “lead with answer, max 2 short paragraphs unless asked.”
Hallucinations stay. Tags make the guesswork visible instead of smooth. You still verify anything that can cost money or reputation.
Workplaces that score “AI usage” still pressure people. Machine mode doesn’t fix politics; padding metrics with nonsense queries is a reported pattern, not a solution this tutorial owns.
Does stripping the voice make the models feel colder and somehow more honest, or does it just move the discomfort somewhere else? I’ve landed on “more usable,” but your mileage may vary after a week of real tasks.
FAQ
Do I still need to use LLMs if I genuinely dislike them?
If you can ship without them, Method A is valid. Most knowledge-work roles can’t fully dodge search summaries and coworker drafts – Kerr’s bar is the practical one there.
Will custom instructions stop all hallucinations?
No. Tone and honesty signals change; next-token guessing does not become a verified database. Money, reputation, citations, API details: treat output as a draft and check it.
What if my company forces Copilot or ChatGPT and tracks usage?
Machine mode still upgrades the quality of what you produce under the mandate. Keep sensitive data out of consumer chats. Prefer enterprise instances with real data controls when they exist. Metrics-as-theater is a workplace politics problem – different from prompt setup, and not something a custom-instructions block resolves.
Open ChatGPT or Claude right now. Paste one block, save, send a prompt whose tone you usually hate. Before/after is the whole tutorial.
Source pieces: Fowler’s original (17 Sep 2026) and Kerr’s essay linked above.