Hi there, and welcome! ★ NEW! ★

My name is Dan Dunford and I am a Senior Software Engineer from Sheffield, United Kingdom. I have been building production software for years — long enough that a page like this one was, at one point, entirely reasonable.

These days I spend most of my time on AI engineering: building the tooling, agents and guard rails that let a language model do real engineering work instead of producing plausible-looking rubbish. I also build the boring parts that make it stick — pipelines, containers, observability, and the kind of tests that fail when they should.

✉ E-Mail Me  |  My GitHub  |  My LinkedIn  |  My Web Log

This is the thing I am most proud of.

Coding agents fail in a predictable way: ask one to "build feature X" and you get drift — inconsistent patterns, skipped edge cases, a confident summary of work that was never actually finished. So I stopped treating the model as a genius and started treating it as a team that needs a process. Then I built the process.

“AI doesn't replace engineering discipline. It amplifies it — if you give it structure.” — me, in one of my write-ups

▸ What it actually is

A private framework that sits on top of my coding agent and gives it three things it doesn't have on its own: procedures (step-by-step workflows it must follow literally, with stop-gates it cannot talk its way past), reusable specs (battle-tested designs for the features every app needs — auth, billing, files, notifications — with a version for each platform I work in), and specialist reviewers (agents with one narrow job, pointed at work somebody else did).

▸ The loop

   IDEA
     |
     v
  [ SPEC ]        argue about what it actually is, before any code
     |
     v
  [ PLAN ]        phases, tasks, and a dependency graph of the work
     |            ................ STOP GATE: plan gets reviewed
     v
  [ BUILD ]       parallel agent teams take independent tasks
     |            each one claims work, builds it, proves it
     v
  [ REVIEW ]      a DIFFERENT model reads it, read-only, and reports
     |            ................ STOP GATE: findings get fixed
     v
  [ AUDIT ]       security / accessibility / mobile / production passes
     |
     v
   SHIP

▸ The rules that make it work

• Whoever builds a thing never signs it off. Review runs on a different model than the one that wrote the code. A tool returning "success" proves the tool ran, not that the work is right.
• Complete or descope — never silently shrink. If a task is too big, it gets broken up or renegotiated out loud, not quietly half-done.
• No placeholders, no TODOs, no commented-out tests. Real code, or stop and ask.
• Replace, don't accumulate. A new implementation deletes the old one in the same commit, callers migrated. This is how agent-built codebases avoid turning into a landfill.
• Verify the running system, not the diff. "Code written" is not done. Did the migration run? Is the data there? Does the page actually render?
• Decisions become checks that fail. An architectural rule nobody enforces is decoration — so the important ones get wired into the build, and each one has to be proven capable of failing.

▸ What it changed

The framework itself is private, but a public spin-off of it — GameClaude, the same process restaffed as a game studio and wired to Unreal Engine 5 — is on my GitHub, along with a series of write-ups on how I work this way.

Things I have actually built and put on the internet. All open source.

No percentage bars. Nobody has ever known what "PHP: 87%" means.

Full CV available on LinkedIn.

Fetched live when you loaded this page. In 1999 this would have been a CGI script.

► See all my repositories

I write about working with AI coding tools — the workflows, the failure modes, and the engineering discipline that has to sit underneath them.

► Read all my posts

Go on. It is tradition. (Entries are saved in your own browser only — nothing is sent anywhere. If you actually want to reach me, use e-mail.)

[email protected]

Available for freelance and consulting work — AI integration and agent tooling, full-stack builds, or infrastructure architecture. If you have a technical problem that needs someone who can work across all of it, get in touch.

▸ Find Me Online