A developer opens a Slack channel to find a wall of text. It is a five-hundred-word response to a simple architectural question, complete with the telltale markers of a large language model: the overly polite introductory paragraph, the redundant bullet points, and the generic concluding summary. The sender did not write this; they simply pasted the raw output of an AI prompt. For the recipient, this is not a helpful answer, but a new chore. They must now spend five minutes parsing the AI's hallucinations and fluff to find the one sentence of actual utility. This scene is playing out across thousands of engineering teams globally as the novelty of generative AI wears off and the friction of AI-driven noise begins to degrade professional productivity.
The Framework for AI-Assisted Curation
The solution to this noise is a shift in how teams treat AI outputs, moving from a delivery model to a curation model. The core principle is simple: never deliver an AI response in its raw form. Instead, the AI should be treated exclusively as a drafting partner. The workflow begins with the user prompting the model to generate a baseline of information, but the process does not end there. The user must then read the output, extract the specific answer to the question at hand, and delete everything else.
In practical terms, this means that when communicating via direct messages, Slack, or Pull Request reviews, the goal is brevity and precision. Rather than pasting a full model output, the standard is to provide three concise, human-edited sentences that summarize the finding. If a specific part of the AI's output is particularly valuable, it should be quoted sparingly and accompanied by a human explanation of why that specific point is relevant to the current project.
Crucially, this protocol addresses the vacuum of opinion. In many AI-mediated conversations, users paste a response to avoid the mental effort of forming a conclusion. The new guide suggests a more honest approach: if the AI provides an answer but the human user has no specific professional disagreement or addition, they should explicitly state that they have no strong opinion. This prevents the illusion of human endorsement when the sender is merely acting as a conduit for a machine.
The Shift from Information Access to Information Judgment
The necessity of this shift stems from the democratization of AI tools. In the early days of LLMs, sharing a complex AI-generated solution provided value because not everyone knew how to prompt the model or had access to the tool. Today, almost every member of a technical team has access to the same models. When the tool is ubiquitous, the value of the raw output drops to zero. If the recipient can generate the same response with a single prompt, the sender who simply copies and pastes that response has contributed nothing to the collaboration.
This creates a fundamental tension between efficiency and laziness. Copy-pasting feels efficient to the sender, but it transfers the cognitive load of filtering to the receiver. By requiring human curation, the burden of synthesis is placed back on the person initiating the communication. The value is no longer found in the ability to find an answer, but in the ability to judge which part of that answer is correct, relevant, and actionable within the specific context of the team's codebase or business logic.
When this etiquette breaks down, the remedy is not necessarily a lecture on professionalism, but a redirection toward established asynchronous communication standards. If a team member continues to flood channels with unedited AI outputs, the guide suggests sharing resources that emphasize the value of the recipient's time. Links to nohello.net or dontasktoask.com serve as subtle but effective reminders that the goal of professional communication is to minimize friction and maximize the signal-to-noise ratio.
As AI becomes an invisible layer in every software tool, the most valuable skill in a collaborator is no longer the ability to operate the AI, but the discipline to edit it.
