Software developer Ashley Childress argues that current spec-driven development practices incorrectly utilize AI as a mere documentation generator rather than a collaborative partner. In standard workflows, developers write a specification document first and hand it to an AI agent to execute, treating the model as a vending machine. This approach often leads to batch thinking, where the model makes a single commitment to a path and follows it through, baking potential errors into the entire plan before the developer has a chance to catch them. According to the article, this method essentially asks the AI to act as a typist, bypassing the critical step of real-time negotiation and collaborative design where core problems—such as edge cases or conflicting requirements—should be identified.

The author suggests that developers should shift their process to prioritize conversational planning. By arguing design choices in chat with an AI configured as an opponent rather than a yes-machine, developers can identify faulty assumptions early. To force this behavior, Childress recommends using a bounded adversarial rule in the model's system prompt. This directive instructs the AI to challenge undecided designs, raise high-risk objections one at a time, and wait for user input before proceeding. This adversarial framework is constrained by stop conditions to prevent the process from becoming exhausting or reopening already settled design decisions.

When this collaborative process is followed, the resulting specification document becomes a byproduct or record of the thinking that already occurred, rather than a stand-in for the thinking itself. This approach is framed as a single-player workflow where one developer owns the design, but Childress notes uncertainty regarding how to scale this conversational model to larger teams where a formal document serves as a shared contract for multiple stakeholders. Ultimately, the article concludes that the value of AI in programming lies in the friction of the conversation, not in the production of a tidy Markdown file. The post itself was generated by iterating through this exact adversarial workflow, where the model was prompted to critique the author's arguments during the writing process.

Software developer Ashley Childress argues that current spec-driven development practices incorrectly utilize AI as a mere documentation generator rather than a collaborative partner. In standard workflows, developers write a specification document first and hand it to an AI agent to execute, treating the model as a vending machine. This approach often leads to batch thinking, where the model makes a single commitment to a path and follows it through, baking potential errors into the entire plan before the developer has a chance to catch them. According to the article, this method essentially asks the AI to act as a typist, bypassing the critical step of real-time negotiation and collaborative design where core problems—such as edge cases or conflicting requirements—should be identified.

The author suggests that developers should shift their process to prioritize conversational planning. By arguing design choices in chat with an AI configured as an opponent rather than a yes-machine, developers can identify faulty assumptions early. To force this behavior, Childress recommends using a bounded adversarial rule in the model's system prompt. This directive instructs the AI to challenge undecided designs, raise high-risk objections one at a time, and wait for user input before proceeding. This adversarial framework is constrained by stop conditions to prevent the process from becoming exhausting or reopening already settled design decisions.

When this collaborative process is followed, the resulting specification document becomes a byproduct or record of the thinking that already occurred, rather than a stand-in for the thinking itself. This approach is framed as a single-player workflow where one developer owns the design, but Childress notes uncertainty regarding how to scale this conversational model to larger teams where a formal document serves as a shared contract for multiple stakeholders. Ultimately, the article concludes that the value of AI in programming lies in the friction of the conversation, not in the production of a tidy Markdown file. The post itself was generated by iterating through this exact adversarial workflow, where the model was prompted to critique the author's arguments during the writing process.