Discover
Understand the business challenge, the operational workflow, the people who'll use it and what they need. Most of the value of a system is decided here, before any code exists.
How I work
Claude Code, Codex and coordinated agents write most of my code. That's a legitimate way to build software when someone is responsible for the requirements, the decisions and the validation. On my projects, that someone is me.
The process
Understand the business challenge, the operational workflow, the people who'll use it and what they need. Most of the value of a system is decided here, before any code exists.
Plan the features, data structures, integrations, automation and user experience, and write down the decisions that must not drift, like "only one module may contact a customer."
Claude Code, Codex and coordinated agents implement the software against the written requirements, one scoped task at a time, with a handoff note at the end of each.
Run the real workflows, find the bugs, verify each integration against the live service, and refine. For AI features, write a test set and measure. A demo that works once isn't evidence.
Release it, run it, watch what users actually do, fix what breaks and keep improving. Two of my systems have run real businesses day to day, so production problems land on me.
Beyond the prompt
Getting an agent to generate code is easy. Getting software that's right, and stays right, takes the skills below. None of them is a prompt.
Knowing the domain well enough to spot when a correct-looking feature is wrong for the people using it.
Turning a business need into data models, rules and tasks small enough for an agent to build and a person to check.
Scoped tasks, written handoffs, decision logs, and one writer at a time so parallel agents don't collide.
Reading what was built against what was asked. Plausible code that passes a glance is where the expensive bugs live.
Tracing a production failure to its cause, such as a job that ran five hours early or an endpoint that 500s with nothing in the log.
Test sets for AI behavior, workflow testing for everything else, then deploying and running it reliably.
In practice
I own requirements, architecture, the agent and system prompts, review, testing, deployment and operation. On the evaluation project that also meant the behavior rules in every prompt version, the held-out questions, and reading every failure myself.
Every project keeps a handoff document and a dated decision log, so an agent picking up the work reads the same constraints I would otherwise repeat. From Galley's log: vendors are shoppers, not warehouses, so no line prices anywhere and spend is the typed receipt total. That's the part a model can't infer from the codebase.
Two agents and a human in one repository will collide. The evals repo carries a baton: a tracked file naming the current holder, with a pre-commit hook that rejects commits from anyone else. Handoff notes are mandatory and written for the next agent.
A prompt is code, but nothing tests it unless you write the test on purpose. My CRM orchestrator still described "a team of sixteen specialist agents" months after the fleet had grown past that. It was true when written, wrong for months, and invisible until someone read it against the code.
I'm open to remote roles at teams that embrace AI-assisted development, and to contract and freelance projects.