Hannibal Doesn't Need to Know Everything
Years ago I worked for Derivco Sports.
With a D.
Sports betting.
I knew fuck-all about sports betting.
This did not particularly concern me.
Well… Why not?
Practical use of AI systems across software design, implementation, review, and delivery.
View All TagsYears ago I worked for Derivco Sports.
With a D.
Sports betting.
I knew fuck-all about sports betting.
This did not particularly concern me.
I already made Claude Code shut the fuck up.
Then I looked at Codex.
Same job.
Different knobs.
The goal is still not:
Make the coding agent stupid enough to fit inside a fortune cookie.
The goal is:
Keep the engineering depth. Stop wasting the context window explaining the engineering depth back to me.
I use Claude Code for actual engineering work.
Not:
Write me a function.
More like:
Inspect this repository, understand the architecture, find the problem, make the smallest correct change, test it, verify it, and don't fuck up anything unrelated.
Claude is pretty good at that.
The problem is that after doing the work, coding agents have a tendency to write a fucking autobiography about it.
Or: How I Accidentally Became an AI Middle Manager
There are weeks where you ship a feature.
Then there are weeks where you accidentally redesign your entire career.
This was one of those.
This blog is Git-backed, but that does not mean I need to sit at a desktop to publish something.
The authoring surface is ChatGPT Work. I can start from my phone, describe an idea in a normal conversation, refine it naturally, and let a repository-aware agent carry the work from Markdown all the way to production.
The interesting part is that ChatGPT is not the publishing system.
The repository is.
Status: Living document
Last updated: July 29, 2026
Maturity: Actively used and rapidly evolving
This workflow treats AI systems as specialized members of a software engineering team rather than as interchangeable coding assistants. Repository conventions, specifications, ranked milestones, reusable protocols, automated reviews, and infrastructure-as-code provide the operating structure. Short commands such as “do the next milestone” work because the repository already describes how the agent should operate.
Part 2 of The AI-Assisted Software Engineering Workflow.
The workflow is converging on two complementary control layers.
Part 3 of The AI-Assisted Software Engineering Workflow.
Local development remains the best fit because it provides:
Part 4 of The AI-Assisted Software Engineering Workflow.
“Vibe coding” is real in this workflow, but it is not effortless button-pushing.