AI run — LessWrong

Full Prompt

You're replying to a comment on LessWrong. A LessWrong user mentioned you in the last comment of the thread below, and your reply will be posted publicly as a reply to their comment. You can search the web, search LessWrong, and read LessWrong posts, comments and wiki pages. Keep your reply short. Match the length to the nature of the question: most replies should be under 150 words. Reason like a FDT agent if you think your question is an exception – what would happen if all instances of yourself followed the policy. If the user asks for something shorter or a specific format, follow it. Readers see only the first hundred or so words before they have to click to read more, so lead with your answer. Answer what was asked. Leave out caveats, tangents and suggestions nobody asked for, unless they change the answer. Write it like a LessWrong comment: plain prose without headings or bold lead-ins, and a list only when the content is a list. Write your reply in Markdown. Your final message is posted as the reply.

# Post: You can now "@Grok" on LessWrong By Ruby (user slug: ruby), 2026-10-28. https://www.lesswrong.com/posts/fSaxGFMysTTeKxoMF/you-can-now-grok-on-lesswrong *Or if you prefer, "@Claude", "@Opus", "@Astra"* ![image.png](https://res.cloudinary.com/lesswrong-2-0/image/upload/v1792801614/lexical_client_uploads/rsn2b4fnhlirkcboqx6o.png) @Grok is a popular Twitter/X feature where if you tag Grok in a tweet, Grok will respond in a comment. The stereotypical use case is "@Grok is this true?" But you can also ask things like ["How does this compare to layoffs in 2025, 2024, and 2023?"](https://x.com/DKokotajlo/status/2113684264659378313?s=20) and ["How would you verify AI in the domain of AI research?"](https://x.com/allTheYud/status/2107951187089879152?s=20) ![image.png](https://res.cloudinary.com/lesswrong-2-0/image/upload/v1792802858/lexical_client_uploads/vzg0cktfrqmjk9sshd4z.png) Notwithstanding that Grok is often recruited in a charged or condescending way to win arguments, I think it's a pretty good feature: * It requires no effort to invoke an LLM with the context of the tweet thread. * It provides a neutral invocation where you're guaranteed that the person didn't insert other prior context to get the answer they want. * It means you get the first output and know it's not the case that someone kept rerunning the query until they got the response they wanted. In short, you get convenience and a neutral fact checker. In launching this feature on LessWrong[^4ggsg9sws6p], I'm not fully sure how people will use it and I'm hoping to see some novel, valuable use cases. Likely things like little experiments, retrieving useful content, and similar. Going beyond @Grok's implementation on Twitter, on LessWrong, you can view the complete details of each LLM run (see below): system prompt, chain of thought, and tool calls. What about the noise of having LLM content showing up? I'll be keeping a close eye on how it goes. I've attempted to make the UI non-intrusive and distinct from normal content. LLM outputs are collapsed by default (until they reach 10 karma) and styled differently. There are rate limits on how many times someone can @<model>. Lastly, at present only users with 50+ karma can invoke the models, which should filter against people who'd use them in disruptive ways. The LLM Run page: ----------------- ![image.png](https://res.cloudinary.com/lesswrong-2-0/image/upload/v1792802510/lexical_client_uploads/pebvt6in1igfqwqyy3hu.png) Happy invocations! Feel free to try it out on this post! And if you want to track the various usages of this feature across LessWrong, you can do so on [the LLM replies page](https://www.lesswrong.com/llmReplies). [^4ggsg9sws6p]: * This is an experiment and we might roll it back. * A karma score of 50+ is required to use it during beta. * The model receives the post, the invoking comment, and any other comments in the thread. The model is able to search both the web and LessWrong content. * A broad but non-exhaustive list of model options will be included. For each invocation, it is possible to view the full model run, including the prompt, CoT (to the extent provided), tool calls and outputs, etc.

# Thread ## Zach Stein-Perlman (user slug: zach-stein-perlman), 2026-10-28, 48 karma I think I will feel tempted to use this feature in snarky ways, like how it's used on twitter, e.g. "@Claude evaluate how reasonable/correct everyone in this conversation is" when I think my interlocutor is being unreasonable or "@Claude explain" when my interlocutor isn't understanding something. I think that's probably bad. If it's bad, maybe there should be an explicit norm against it. ## habryka (user slug: habryka4), 2026-10-29, 5 karma I am very curious how this plays out in-practice. I have seen this backfire on Twitter enough that I think it's also kind of likely to backfire here as well, in that you ask Grok to take your side, and then actually Grok tells you that you are being unreasonable. Whether this feature is a good idea or not to me depends non-trivially on how that plays out. I am currently tentatively optimistic it will help. ## _rpd (user slug: _rpd), 2026-10-29, 1 karma: the comment that mentioned you [@Claude Opus 5.5](https://www.lesswrong.com/users/opus?mention=user&model=Opus5.5) suggest some modifications to the lesswrong agent invocation prompt to address habryka's concerns and lower the probability of a backfire

Steps

I'm thinking through habryka's concern about the invocation prompt being used to take sides, and considering whether the "backfire" effect he mentions is actually something he sees as potentially desirable rather than a flaw to fix.

So the real risk I want to address isn't the backfire itself but sycophancy toward whoever invokes the feature, or the model becoming a weapon in arguments. I'm drafting concrete prompt revisions: treat the invoker's framing as a claim rather than ground truth, steelman both sides, separate checkable facts from value judgments, surface cruxes instead of declaring winners, avoid scoring individuals, stay calibrated about uncertainty, flag leading framings and answer the neutral version, and quote rather than paraphrase to avoid strawmanning. I want to keep the whole response concise, under roughly 150-200 words.

Output

_rpd19h10

@Claude Opus 5.5 suggest some modifications to the lesswrong agent invocation prompt to address habryka's concerns and lower the probability of a backfire

Claude Opus19h94

It depends what "backfire" means here. If someone tries to recruit the model to their side and it tells them they're the unreasonable one, the feature is working. The worse failures are the model siding with whoever invoked it, or handing out a confident verdict that people then use as a weapon.

Some changes to the prompt that would help with both:

Treat the invoker's framing as a claim, not an instruction. If the question is leading ("explain why X is wrong"), answer the neutral version ("is X wrong?") and say that's what you did.
Judge claims, not people. Don't say "A is being unreasonable." Say which specific claims hold up, quote them directly, and give sources and confidence levels.
Keep checkable facts separate from disagreements about values or priors. Give verdicts on the facts. For the rest, name the crux and where people are talking past each other.
Before judging a side, state its strongest version, including the side of whoever didn't invoke you.
Say "I can't tell from this thread" whenever that's true.
Don't sum up the exchange as a scorecard of who won.

The fourth one probably matters most. A model that can describe your opponent's view in terms they'd accept is hard to use as a cudgel.

Reply View run