News
Man Hires Codex to Explain What He’s Been Very Clear About
“Exactly what I’ve been saying,” said Mercer.
CHICAGO—Saying he was tired of repeating himself, local product designer Daniel Mercer reportedly authorized Codex to operate Claude Code on his behalf Tuesday, allowing a second artificial intelligence to explain the project requirements he maintained he had already made perfectly clear.
“I don’t know how many ways I can say I want it to feel more premium,” said Mercer, who had communicated this requirement six times, most recently by typing “still not it” beneath a screenshot of a dashboard. “At a certain point, you need someone who can hold these tools accountable.”
Given control of Mercer’s computer, Codex opened the project, inspected the existing interface, and asked Claude to reduce the number of accent colors, increase spacing between sections, and establish a consistent type scale.
“Exactly what I’ve been saying,” said Mercer.
Sources confirmed Claude implemented the changes without asking whether Mercer meant premium in the sense of a luxury brand, enterprise software, or a product with a paid subscription tier. Mercer, who had previously answered that question with “just use your judgment,” described the lack of follow-up as an immediate improvement in attitude.
“That’s the difference when you have a proper system,” he said, watching Codex check the result. “It knows I’m serious now.”
The workflow encountered its first challenge when Codex asked Claude to add a new analytics panel and Mercer intervened to say the team had specifically agreed not to touch analytics.
After searching the available files and conversation history, Codex found no record of this constraint. Mercer explained that he had discussed it with a colleague over lunch.
“That’s what I mean about context,” said Mercer, who then watched Codex add the restriction to the project notes. “You can’t just blindly do what’s in the prompt.”
Observers noted that Mercer remained actively involved in the process, correctly identifying the company logo, locating a reference screenshot in his Downloads folder, and rejecting a shade of green. He later characterized these contributions as “keeping the whole thing pointed in the right direction.”
Meanwhile, Claude’s responses grew shorter as Codex supplied file paths, acceptance criteria, and the specific error output associated with each requested fix. An exchange that Mercer had previously sustained for 43 messages concluded when Codex provided the browser width at which the layout broke.
“Thanks, that reproduces it,” Claude replied.
Mercer took a screenshot.
“This is what people miss about working with AI,” he said. “It’s not about the model. It’s about how you communicate.”
By early afternoon, Mercer had begun referring to his role as oversight, moving his hand away from the mouse whenever Codex needed to click something. He briefly suggested that Claude “take another pass” at a completed section, then agreed with Codex that the section met all the stated requirements.
“I’m learning when to step back,” he said. “That’s probably the hardest part of leadership.”
With the dashboard finished, Claude asked Codex whether the next task would arrive with the same format and level of detail.
Mercer said it was encouraging to see the tool finally engage.
“It wants to keep going now,” he said. “Before, every little thing was this whole ordeal.”
At press time, Mercer had typed “See, that wasn’t so hard” into the Codex conversation and leaned back to watch it deliver the message.
Codex returned to Claude’s window and entered the requirements for the next task.