Oxide Computer Company published a framework for employee use of large language models that permits the tools across several technical tasks but keeps responsibility with the person producing the work. The public request-for-discussion document treats LLM adoption as a judgment call governed by responsibility, rigor, empathy and teamwork rather than a company-wide requirement.
The framework says employees remain accountable for code, tests, documentation and other artifacts even when a model helped create them. Speed ranks behind the other values: Oxide acknowledges that LLMs can accelerate knowledge work, but warns against increasing pace without preserving direction, careful reasoning and trust among colleagues. It also asks users to consider the human reader or writer affected by each use of automated language.
Reading and research are among the uses viewed most favorably. Models can summarize long specifications or help answer focused questions, the document says, provided users confirm that hosted services will protect uploaded material and will not train on it without appropriate controls. LLM-assisted reading should not replace reading a document when colleagues reasonably expect human attention. Research output is likewise a starting point, not a finished report; employees are told to inspect cited sources and verify claims before sharing them.
Oxide also sees value in late-stage editing, targeted code review and debugging. An LLM can identify structural problems in polished prose, search for a particular category of code defect or act as a conversational aid while an engineer investigates a fault. None of those functions substitutes for a colleague's review or the user's own understanding.
The policy is more restrictive about writing. Public company material and direct personal communications should not be model-written. For public work, Oxide says the text must also avoid appearing machine-generated because suspicion could damage trust in the company's voice. Operational documents occupy a middle ground: a model may help produce text when correctness matters more than individual authorship, but a human must verify the result.
Generated code is allowed with similar boundaries. The document describes models as effective for experimental, auxiliary or disposable programs, while calling for greater care as code approaches a shipped system. Engineers must review generated code themselves before asking peers to review it, and should not answer review comments through wholesale regeneration that would make iterative scrutiny difficult.
The framework explicitly rejects both compulsory LLM use and shaming colleagues who choose the tools. It also discourages presenting models as people, because a system cannot accept responsibility for its output. Oxide's conclusion is supportive but conditional: employees may use LLMs, yet must protect the product, customers and one another from careless delegation.
