A four-hour Claude Code session landed 42 unreviewed commits on main

Step three is where Souza spends his effort. He hands over context files written in the company's own language, asks the agent to read everything and report back what it understood without touching code, and only then asks which questions it wants answered before planning starts ^[12][13][14]. The agent writes the questions into a file. He answers them one at a time ^[14].

Some questions he did not understand, so he asked for a better explanation. Others made him realise he did not understand the module properly himself ^[15]. That produces a second round, then a third and a fourth ^[15]. The convergence signal he uses is the scrollbar on the question file shrinking with each round ^[16]. When the doubts run out, Souza wrote, he sends a short prompt, "pode seguir até o final da implementação, confio em você", roughly go on to the end of the implementation, I trust you, and leaves it running ^[17]. The rounds are not his invention; he wrote that Anthropic itself recommends that kind of exchange with the agent ^[18].

Forty-two commits in a little over four hours is about ten an hour, one every six minutes ^[1]. Counted with the smaller cases, the module was fifteen or sixteen use cases ^[10], so roughly two and a half commits each ^[2]. Souza's own correction to all of it was a single adjustment in one use case ^[3].

For the four hours to mean anything on another team's repository, the starting conditions have to match. This module was the second part of the system, and the first part's ten use cases Souza built one at a time ^[9]. Fourteen iterations of context were already in the folder when he gave the final brief, and he says that last iteration concentrated a lot of context ^[4][11]. On an empty repository, the context files would have to carry the conventions that those ten hand-built use cases carry here.

The test counts are what I would want explained next. The backend finished the session with just over 1300 tests and the agent with just over 800 ^[6], more than 2100 between them ^[3]. Souza does not say who wrote them; he says he will come back to them later because they have a curious role in the story ^[7]. The post itself was initially conceived by generative AI from the transcript of his Dev Eficiente video ^[22], which is at least consistent.

Souza's frame is management. He led teams with a technical bias and could do, at some scale, what the people on them were doing ^[24], and he wrote that from his point of view the agent is the team that will execute the work ^[19]. He also wrote that he does not think there is a right prompt or a playbook for Claude Code, Codex or any tool of the kind, and that he reads the documentation of the tool he uses most plus what Anthropic and OpenAI publish, and forms his own opinion ^[20]. That opinion, in his words: "dá contexto, tira as dúvidas, solta para o jogo" ^[21].