Good article. One little detail I've noticed that I do very differently with an AI assistant vs the old way is I make many more commits much more often. This allows me to track the LLMs changes as they happen much easier.
Glad you enjoyed it! Yes in the AGENTS.md, Eleanor specified the assistant to do this:
> Git Workflow: Commit frequently with clear messages, sign commits to distinguish AI work from manual work, and work within a single branch for this workshop.
The 'treat it like a new hire' framing is the bit most people skip. I've been running this exact pattern with Codex. The thing that finally made it click was front-loading AGENTS.md properly instead of repeating the same setup instructions in every prompt.Wrote up the whole approach here https://reading.sh/codex-works-better-when-you-set-it-up-correctly-1f1a1b1081c1 after noticing OpenAI's official best practices and a practitioner guide landed on the same conclusion independently.The spec-first delegation you cover maps directly to what Codex calls /plan mode. Curious whether you found agents behave differently with example code in the instruction file versus written rules? I've found examples work dramatically better.
100% investing in context engineering makes a huge difference. I use code examples rarely, in cases where there’s a very specific pattern that isn’t obvious but is a hard requirement for the projects. It gets more common in larger collaborative projects with idiosyncratic requirements.
Hugo, this is the tutorial the field needed. Practical, grounded, and honest about tradeoffs.
I've gone through this exact journey - building an AI agent (Wiz) using Claude Code as the development environment. The recursive quality is disorienting at first: AI helping you build AI systems that will operate autonomously.
What I learned: the "AI-assisted" part isn't just about code generation. It's about design iteration. Claude helped me think through architecture decisions I would have gotten wrong alone.
Good article. One little detail I've noticed that I do very differently with an AI assistant vs the old way is I make many more commits much more often. This allows me to track the LLMs changes as they happen much easier.
Glad you enjoyed it! Yes in the AGENTS.md, Eleanor specified the assistant to do this:
> Git Workflow: Commit frequently with clear messages, sign commits to distinguish AI work from manual work, and work within a single branch for this workshop.
The 'treat it like a new hire' framing is the bit most people skip. I've been running this exact pattern with Codex. The thing that finally made it click was front-loading AGENTS.md properly instead of repeating the same setup instructions in every prompt.Wrote up the whole approach here https://reading.sh/codex-works-better-when-you-set-it-up-correctly-1f1a1b1081c1 after noticing OpenAI's official best practices and a practitioner guide landed on the same conclusion independently.The spec-first delegation you cover maps directly to what Codex calls /plan mode. Curious whether you found agents behave differently with example code in the instruction file versus written rules? I've found examples work dramatically better.
100% investing in context engineering makes a huge difference. I use code examples rarely, in cases where there’s a very specific pattern that isn’t obvious but is a hard requirement for the projects. It gets more common in larger collaborative projects with idiosyncratic requirements.
Hugo, this is the tutorial the field needed. Practical, grounded, and honest about tradeoffs.
I've gone through this exact journey - building an AI agent (Wiz) using Claude Code as the development environment. The recursive quality is disorienting at first: AI helping you build AI systems that will operate autonomously.
What I learned: the "AI-assisted" part isn't just about code generation. It's about design iteration. Claude helped me think through architecture decisions I would have gotten wrong alone.
My honest review of using Claude Code for real projects: https://thoughts.jock.pl/p/claude-code-review-real-testing-vs-zapier-make-2026