I’ve spent 25 years shipping software, including the messy brownfield kind. The kind of codebase with a database older than some of its developers, custom types nobody remembers naming, business logic spread across layers that have outlived three rewrites, and a UI that’s been through more frameworks than I can count.
That’s the codebase most working engineers actually live in. It’s not the codebase most AI demos are built around.
This publication is field notes from inside that gap.
Across decades of shipping real software, I use AI as the primary tool today: orchestration, prototype pipelines, locator agents, hook-enforced TDD, handoff files as inter-session memory. The methodology is refined daily, against real codebases, with real constraints. None of it is theoretical. All of it survives contact with code that was written before “AI coding assistant” was a category.
Who this is for
- Engineering leaders and CTOs trying to figure out whether AI methodologies survive contact with production systems that aren’t greenfield React apps.
- Senior developers and AI-tool builders who’ve hit the orchestration wall — where adding more agents stops helping because context discipline is the actual bottleneck.
- Engineering managers and product leaders evaluating whether delivery speed claims hold up when the codebase fights back.
Who this isn’t for
- Greenfield-demo enthusiasts. Most of what I write is about codebases that already exist and can’t be paused.
- Anyone looking for AI-tool reviews. I write about systems and disciplines, not which model is current this week.
What you’ll get
Specific methodology. Workflow command structures, agent definitions in the abstract, the kinds of hook scripts that enforce disciplines automatically, the tradeoffs that come with each one. The patterns that failed before the ones that worked.
The examples are anonymized — never tied to any specific client, product, or codebase — but the methodologies are real. They come from work I’ve done on real systems over a long career.
Cadence
Bi-weekly when the workload permits. Posts are 1,500–4,000 words depending on whether they’re punchy techniques or flagship deep-dives. Comments on, subscribers-only.
If you’re working on real production code with AI in the loop — subscribe. The next post will be useful.
— Adam
