Llewellyn Falco

AI works well on greenfield code and falls apart on the code most of us actually have. In this hands-on mob session, we take a real legacy codebase and make it safe for AI, live, in small steps and fast feedback loops.

Together we write the handful of CLIs every project needs: a fast test script, a coverage script, and format and lint tools. We wire in Habit Hooks, which turns linter findings into coaching an agent can act on, and use it to take on smells like duplication and long methods. We use deliberately fast, low-thinking models, because tiny steps and quick verification beat clever prompts.

Then we extract knowledge into documents: product docs, process docs, and agent rules. We clear context repeatedly and reload only what the agent needs from those docs, so the context is something we build and keep rather than something we re-explain every session.

**You will leave able to:**

• Build the basic CLIs a legacy project needs before an agent touches it
• Wire in hooks that turn deterministic checks into guidance for the agent
• Extract product and process knowledge into docs and rules the agent can reload after every context clear