Process

How these were built

I directed these projects: ChatGPT and Claude helped design and plan them, and Codex and Claude Code did the technical work. This page sets out who decided what, how the results were checked, and what that means when you read the work.

Who did what

I decided

  • The business context, the purpose of each project and an initial architecture outline
  • Requirements and evaluation criteria
  • Which edge cases mattered most
  • The judgment calls, and whether each result was approved

AI tools did

  • ChatGPT and Claude: design and planning, and generating the agent prompts
  • Codex and Claude Code: the code, tests and documentation, including the B2B case study’s documents, fictional scenario and checking scripts
  • Codex and Claude Code: verifying later corrections and their regression results

From brief to published project

  1. Write the brief

    I co-developed the original agent prompt kit with ChatGPT and Claude. For each project I provided the business context, the purpose and an initial architecture outline, and set the requirements, evaluation criteria and edge-case priorities.

  2. Plan, then build

    A planner-and-builder AI workflow turned the brief into code, tests and documentation, one project per session, with Codex and Claude Code doing the technical work.

  3. Build the checks in

    Every repo includes checks anyone can run: 21 tests for data quality, 130 for the inventory model, 44 for analytics (including headless dashboard runs), and 30 tests, a scenario consistency check and a document checker for the B2B case study.

  4. Review in a fresh session

    Before publication, each project was reviewed in a fresh AI session. In the first review round I also followed each project’s run-and-check instructions in PowerShell myself and confirmed that the outputs matched the agents’ reports. Later corrections and their regression results were verified by agents.

  5. Publish by hand

    Nothing goes live automatically. I review the changes and publish them myself.

Boundaries

  • The projects live on a separate, personal GitHub account.
  • The AI tools were given no data, code, or proprietary documents from my job.
  • All data is fictional. Each project is a simplified, independent version of a problem related to my work, not a copy of it.

This site

ChatGPT built the first version of this site. Claude then redesigned and restructured it, writing the code and drafting the copy on branches under my direction and opening pull requests; I made the judgment calls and merged them.

Scripts in the repo generate every figure from numbers copied from the project repos, enforce minimum label sizes and contrast, and check links and placeholders. Each redesign phase was also fact-checked against the repositories by a separate Claude session before its pull request was opened.

What this means for reading the work

  • The code shows what I specified and accepted, not code I typed by hand.
  • The case studies focus on the decisions: the requirements and edge-case priorities I set, and the trade-offs I approved.
  • The tests let you rerun the checks yourself rather than take my word or the AI’s for the results. They show the code does what the tests expect, not that it is right in every case.
See the projects