Stop Being a 'Meat Proxy' for Your AI

A common workplace pattern has emerged where people paste a question into Slack, a pull request, or a group chat and reply only with a raw, verbatim block of Claude output. The author argues this adds no value: anyone can query an AI directly, and doing so is faster and gives them control over the context. Relaying unedited AI text just inserts a human middleman — a ‘meat proxy’ — between the recipient and the model.

The deeper problem is that AI output is costly to consume. It tends to be verbose, dense with jargon, and salted with plausible-sounding nonsense that still has to be verified. The author cites a single Claude-generated sentence about NATS quorum re-formation that required looking up nearly every term. The value a person actually contributes is the work of reading, understanding, and validating the output, then rewriting it in their own words — which doubles as evidence that they did that vetting.

Code review is offered as the sharp example. It’s now possible to ship code with almost no engagement: paste a ticket into Claude Code, don’t read the result, funnel any reviewer feedback straight back in, and iterate. It works, but it hollows out accountability — the reviewers become the de facto implementers via the tool, while the nominal author contributes nothing but relay. The piece is a call for human judgment to stay in the loop rather than being reduced to a passthrough.