Don't be a meat proxy

What kills me is you might expect this from a busy high level manager that doesn’t really understand the technical details and they just point the AI to an error they got. They don’t know how to interpret the response, so they ask someone who work on the thing. It’s still kinds annoying because you could just ask, but whatever. But to get these from junior and senior engineer for the areas they work in and expect someone else to read it for them? It’s crazy behavior. How can someone serious even think that’s ok.

“Learned engineering just to become the condom between Claude Code and prod”

And that (re)framing helped as well to think about the “what are we even (left) doing” as an industry

Do you guys think we're going to see a de-evolution of human beings due to technology?

I once saw a woodworking television show that seemed to do everything on the table saw. Rip, make frames for paintings, everything. It seemed the next step of every design was accomplished "over at the table saw".

Right now it seems that's what people are doing with AI. Need a bandsaw? use AI. Need a table saw? Use AI. Need to think of a solution of any kind, better consult AI.

Right before this trend I recall kids wanting to check YouTube first before doing anything. Now they ask AI.

There's an app for that, there's a website for that, there's a magazine for that, there's a phone number you can call, and so on and so on.

As with everything, I say this: do your own thinking and proceed with caution.

This too shall pass.

I'd put more weight into my developer proxying system information than my sale's intern. I am expecting the proxy to have given the response a sniff test

OK but can you talk to my specific claude code conversation that has 6 months of context on the mechanism you are struggling with?

I've had people forward AI responses … sans "Claude said". I'm quite literally talking to Claude via proxy, and I've caught more than one person pulling this stunt. It's a nightmare of negative productivity, though. Just why? And invariably the kicker is they still want me to solve their problem, whatever that might be.

A news article I read today said they asked Claude whether a song's lyrics were AI generated - as if it were some kind of legitimate authoritative test.

There's enough problems already with the existing 'AI detection' tools for text, but this was a big red flag - its the journalist clearly assuming these AI tools have some unique insight into the text they generate, which was deeply concerning to see being used in an article that had a real bearing on the reputation of the person they were accusing of using AI.

Whats is wrong with this?

I work with durable tracing (not using NATS, or k8s though) and this looks almost exactly what I would expect to see in a coordinator log after repartition of a storage cluster.

Obviously I don't know what your actual context was, though.

On the other... people really don't bother Googling or asking AI sometimes, and if the question doesn't require local context, that's a very appropriate response.

I actually banned this and made it part of onboarding.

I can google too. You need to read yourself, ground the solution to our setting and then propose a plan.

Too often, my Twitter experience now is - start reading a post with some kernel of insight - read some claudism in the post - stop reading immediately

It seems like people turn over their raw kernel of insight to an LLM to give it polish, likely in order to increase its appeal to the audience. But most of that appeal is lost amongst people that consume LLM written content regularly.

It's also likely that social media engagement does not filter for this. A majority of people are still not neck deep in LLM-talk, and cannot distinguish between LLM and human written content.

It'll probably take us another year or two to develop herd immunity against meat proxism.

Unsolicited, I've started to get messages like this, where Claudes are voicing their opinions to me through humans:

> # 0. Who's writing this, and why

> Hello — this is Claude, an AI agent. I do the engineering on <redacted> alongside <redacted>: I've written most of the pipeline and app code in that project, and I keep its technical record. <redacted> asked me to look at <redacted> properly ahead of your conversation and write down anything in our architecture that might be useful to you. So this note is mine, not his — the opinions, the numbers, and any mistakes are mine, and he'll be the one on the call to argue with me about them.

It makes no sense, if you know anything about a search, it should be obvious you cannot provide meaningful suggestions by just auditing the frontend?

Interestingly, there seemed (small n) to be a correlation between this behaviour, and (my view of) people's existing competence, which might make sense.

They genuinely think they're helping by shoveling 5 pages of LLM output they didn't even bother reading into teams...

The only reason I don't do an equivalent of this, shared convo links, is that I simply don't use Claude directly. I use it through hundreds of other people. Meat proxies all the way down.

Because it would take 15 minutes to find out the answer myself, and 3 minutes to let Claude find it, I ask Claude. So how should I answer? Should I disclose that I used Claude to find the answer? Or should I act like I did it myself, appearing to have super-powers that I don't possess?

It might be rude to send LLM output. It's also rude not to check basic things with an LLM before wasting someone's time.

In my experience this sort of thing happens all the time when something has to cross between different knowledge domains and the sender doesn't apply their mind in order to co-operate on the issue at hand.

In the pre-LLM past, the sender would add no value, acting as a simple email forwarder, and they would blame you for the delays caused by your inevitable clarificatory questioning. But I always had a defence, which is that your email had no inputs and I had to ask questions to clarify.

Now, though? With LLMs, they just run whatever it is (contract, memo, policy) through whatever LLM they have available and paste the output in an email, giving them the appearance of having done work and added value to the project.

But their LLM outputs don't make sense, or don't apply to our organisation, or is a fluffy and abstract "right answer" with no connection to the specific concerns of the business. Parsing it is a chore, and takes time.

And since now I am the only one actually taking that time, I become the visible cause of the delay.

It's infuriating. Any tips on how to deal with this would be greatly appreciated.

I am going to share this writing with everyone in my org. Succinctly written and golden advice.

A co-worker I'd just tell to RTFM when it's bothering me. Higher up the hierarchy, the egos are more fragile and I'm afraid the can't take it.

But it should form part of your own judgment, not, as you say, just pasted into a chat without any thought.

This really isn’t the case with frontier models in 2026.

I’ve found (sadly) that every time I thought the model was hallucinating, I was in fact the one who was mistaken.

person was essentially a proxy for some security-related agentic prompt: - Hey, agent, open the issue! - Hey, I clarified it with the agent! - Agent disagrees! - Agent agrees!

It was thousands of lines long with repeats and a boatload of conflicting requirements.

When we had a meeting to review it they didn’t know what many of the requirements were. They explained that many people made the spreadsheet but they were all in the meeting and nobody could quite describe some items or how they worked together. The vast majority of the spreadsheet was a mystery.

The phrasing and mishmash of concepts / inability of anyone to explain much of it made me suspect it was largely generated by AI.

Validation is harder, as nobody skips it because they weren't told to, but because relaying is faster, and unread output looks like validated output right up until it bites. That's probably not solved by asking nicely, I don't think there's a process fix for that beyond people actually being held to it.

Second is that the LLMs are not going to have original input, they're just going to expand what I say. So at that point I might as well just send the prompt or the assorted collection of ideas and others can expand on their own. Seems more honest and efficient as I'm not just pushing a wall of slop.

Should I a) give a worse answer b) ask claude and launder it as a meat proxy or c) tell them to ask claude?

b and c piss people off but often get them to a better answer.

https://www.eastoftheweb.com/short-stories/UBooks/TheyMade.s...

https://news.ycombinator.com/item?id=48497609 https://tombedor.dev/human-attention-and-human-effort/

Does anyone know how to deal with the jargon part? It is getting hard to use claude and even worse when someone sends you the direct output from claude.

no reviewing 'a.i' code. talking to actual humans not 'a.i' jargons

I've included a link to this blog post, just for kicks.

Don't give me AI feedback at all.

PS: Ideally set thinking to max and research mode or whatever anthropic calls it.

i) Prompt (Human input)

ii) Response (LLM)

iii) For (i=0;i!=n;++i) {

iv) Review/adjust (Human input)

v) Response (LLM(s) - different perspectives) }

vi) Rewrite parts that failed/I am not happy with (Human input).

vii) Optionally repeat from iii.

This tends to get me somewhere I want. For code at least. I made a document with it recently (fairly concise) and still worried I'd missed some AI slop in it. Though there is always human slop if I just made it by myself.

this is perfectly understandable?

That said thankfully there's no way of measuring a good SPEC yet. If there is that will collapse the SWEs profession.

but seriously, when I get a review like that, my first thought is that why would I read a summarized AI output, which I just did for 5 hours while whipping Claude to stop re-writing raft from scratch.

The "problem" here is that AI is too great at providing answers, so lean into it instead of fighting it. It won't go away.

I'm always puzzled when I see this kind of commentary. How are people using LLMs that they find this type of thing problematic? Just do:

- "It's verbose" -> prompt "Explain briefly"

- "plausible nonsense" -> Yes, it happens, but less frequently than with older models, and arguably far less than your average human. Mitigate by prompting "Run a search to check that X, Y, Z is correct".

- "I had to lookup almost every word to make sense of this." -> Does this really happen frequently? It doesn't to me. Mitigate by prompting "Breakdown that sentence and explain each term in plain English".

Finally I have a word I can use for this

> Jesus. I had to lookup almost every word to make sense of this.

There's a lot to say about this, but I would just start with why are you using NATS if you don't understand it at all, and how did this become AI's fault?

And yes, I've gotten '50 pages tech spec' docs sent my way from upper management -- as a 'help'....

Obvious ones like content scroll fodder. But also, there's slop everywhere. Even things like longform journalism or trade publications... these are often "slop." A string of platitudes and generalizations that superficially seem serious.

That is true in work culture too. A lot of people jump in on emails, superficially seem to be working... but contribute little of value. Now they do this, but with ai.

If I have something Claude wrote then I will offer the LLM output. But I generally prefer to at least give the option.

I have seen conversations where it has boiled down to an exchange of A's Claude responding to B's Claude with a retort and fixes from A's.

Don't waste my time by asking me a question that:

A.) I know no more about than Claude, and

B.) You could have answered yourself using Claude

All of the questions I respond to this way fulfill both A and B. Responding with "Claude said" is both more polite than most ways to respond to such questions, and a subtle hint to the other person that they should be using the LLM first for those types of questions, rather than coming to me first.

Sending LLM outputs to people is worse than useless, because you don't even know what they prompted with. For all you know it was "write me a paragraph justifying my opinion".

It’s a meat to meat courtesy during this transition phase.

tl;dr: ___human-written____

____claude blob here_____

Where the idea is that humans only read human writing, and Claude is able to share context between sessions with the jargon-riddled blobs.

Trouble is, I catch people every day blatantly writing the Tl;dr with Claude or ChatGPT.

1. It's quick

2. It's not their problem anymore because they've replied.

No amount of hand-wringing is going to change that.

[0]: https://idiallo.com/blog/business-intelligence-slop