The post coins “meat proxy” for a now-common habit: someone asks a human a question, that human asks Claude or another model, then forwards the raw answer back without checking, rewriting, or owning it. The complaint is not “don’t use AI.” It is that raw model output is expensive for everyone downstream. It is verbose, jargon-heavy, often detached from local context, and cheap to generate but costly to verify. The core objection landed hard with people dealing with AI-written PRs, docs, Slack replies, and even meeting summaries that nobody seems to have actually read.
The strongest pattern running through the comments is cost shifting. People kept coming back to the same asymmetry. One person can generate thousands of lines of “documentation” or a giant diagnostic writeup in minutes, then dump hours of review work onto teammates. Several framed this as
Brandolini’s Law applied to LLMs, or as a gap between generation capacity and verification capacity. That framing stuck because it explains why even technically correct AI output can still be anti-productive. The problem is not just hallucinations. It is forcing others to parse dense prose, reconstruct missing context, and decide what matters.
Most people were not rejecting AI use outright. They were drawing a line around accountability. Use Claude to research, debug, draft, or explore unfamiliar systems. Fine. But once the output is headed to another person, the sender should have read it, compressed it, translated it into local language, and be able to defend it. A recurring standard was simple: if you still feel the need to say “Claude said,” you probably have not processed it enough. Several teams now ignore raw AI dumps, require a human
TLDR, or push people to summarize findings in their own words before asking for review.
A second thread pushed the blame upward. Many said this behavior is being rewarded by managers and executives under pressure to become “AI-native.” That shows up as AI-generated interview feedback, bloated documentation mandates, managers pasting chatbot answers into technical conversations, and orgs treating visible AI activity as productivity. This is why many commenters did not see the issue as a few lazy engineers. They saw it as incentive design. If leadership rewards output volume and “using AI” more than understanding, people will naturally optimize for buck-passing.
There was also a noticeable undercurrent of dread from experienced engineers who feel their work has shifted from building systems to babysitting agents and reviewing machine-generated churn. Some described real productivity gains in narrow cases. Others said the job is becoming less intellectually satisfying even when throughput rises. Still, the more grounded conclusion was not that software engineering is over. It was that the scarce skill is moving toward context management, judgment, and review discipline. Teams that keep those human functions intact seem to get leverage from LLMs. Teams that let them atrophy get slop, hidden labor, and brittle systems.
gruhn.me
The Advantage and Disadvantage of History for Life