Meat proxy

A human who relays AI output without reading, understanding, or validating it — the person becomes the I/O device; the model does the thinking. The inverse of vibecoding: in vibecoding the human drives and the model types; in meat-proxying the model drives and the human carries.

Origin and meaning

ngruhn, “Don’t be a meat proxy” (captured; HN 49151933, 1,849 points / 741 comments — captured in full). The essay’s trigger: a Claude sentence so jargon-dense (“NATS control-plane events: stream leader election / R3 quorum re-form during pod churn”) the author “had to look up almost every word.” The diagnosis: “Reading AI output is extra effort. It’s verbose, frequently contains all too plausible nonsense, and is increasingly jargon dense.” The rule: “By all means, prompt AI. But don’t just relay the output. Read it, understand it, validate it, and then write a response in your own words (a decent certificate that you’ve done the prior steps). Making that effort is value you can add.”

The code-review scenario that named the pattern: ship code by pasting the ticket into Claude Code, never reading it, pasting reviewer feedback back in too — “That works. But who has done the implementation? The reviewers did… and you as a meat proxy.”

The thread’s field notes (all in the captured HN page)

• “It’s our absolute obligation to make ‘meat proxy’ stick as hard as ‘vibe coding’ did” (afisxisto) — the term’s memetic ambition, explicitly stated
• “Remove meat proxy before flight” (tmcnally) — the pre-flight-checklist joke; the shortest statement of the cure
• “Claude told me to tell you, you’re right” / “Claude says this is funny” — the thread’s recursive punchlines: meat-proxying about meat-proxying
• The cynical corollary (thenatureboy): “I know a guy who just started and uses LLMs all day to submit his work product. He’s getting ahead… people are going to be meat proxies and be promoted over you. Whining on HN hasn’t changed this dynamic and it never will.”
• The surrender confession (Malcolmlisk): quality-focused engineer watches slop-shippers get promoted — “now I just let opencode do my work while I watch and play dota in the background”
• The defense: “‘Claude said’ is honest; worse is people sending slop as their own work” — attribution as the least-honesty bar
• The cost accounting (jdalsgaard): 20 pages of docs in 30 min looks good for you — “having 100s of people read through that slop will cost the organisation… in lost trust, respect, joy of work, sense of purpose”
• The pre-slop precedent (netcan): “when AI slop became a concept, it mostly came into spaces where human slop had already been firmly established”
• The counter-thesis (smolder): “There is no meat proxy subverting discourse. There are just a bunch of dumb disinterested people in the way of a real community.”
• Related coinage in thread: “slop grenade” (pasting AI output into a chat), “BI slop”

The memes (2026-10-06)

• MarkPieszak (854❤ / 37k views) — the Gordon Ramsay “idiot sandwich” parody, vision-transcribed into the clipping: two raw ribeyes pressed to a head, “WHAT ARE YOU” / “MEAT PROXY”. Caption: “imagine having decades of engineering experience… and this is what it all comes to now.” The joke’s precision: she’s not the meat — she’s the proxy for the meat; the vehicle presenting it.
• The term’s grammar generalizes: “meat proxy” as prefix — s/meat proxy before flight, “too many damn people are meat proxy” (AndreiCalazans, in-thread).

Where the vault disagrees with the cynics

The promotion-watching surrender (meat-proxying pays, quality doesn’t) is the honest read of this quarter’s incentives — but it’s the exact dynamic Vibecoding vs AI engineering maps: the market pays output until verification fails, then pays verified output. Uncle Bob’s compiler analogy (Software engineering in AI era) holds: nobody reads the binary, but somebody owns the structure. The meat proxy owns nothing. Spec Gate, explain-then-fix, and the eval stack (LLM evals) are the anti-meat-proxy tooling: they make “I didn’t read it” detectable.

Vault kin

• Vibecoding vs AI engineering — meat proxy = vibecoding’s degenerate terminal state (human as transport, not driver)
• Software engineering in AI era — the responsibility question from the continuity side
• LLM evals / trajectory graders — mechanical detection of relay-without-understanding
• jnsahaj’s explain-then-fix (“reject non-understanding”) — the skill-level cure for the code-review flavor
• explain-in-your-own-words as “certificate” — same test as Feynman technique / the socratic lane

Practice

• Never relay AI output you couldn’t defend line-by-line if the model were down.
• Own-words summary at the top of any AI-assisted message (clbrmbr’s org policy: human tl;dr over the Claude blob — and yes, “I catch people writing the tl;dr with Claude”).
• If you catch yourself pasting between two agents without reading either side — you’re the proxy; remove yourself from the loop or engage.