Writing

Essays.

I write rarely and only when I have something definitive to say. Each of these is researched, grounded in real recovery work, and meant to settle a question rather than circle it. No hot takes. No content cadence.

Recovery · Governance

The 20-Page Termination Deck

Half of all technology programmes are late, over budget, or cancelled. The UK spent £12.7 billion on a single IT programme before walking away. My worst engagement produced five days of delivery in two months. A termination deck is the clearest brief you will ever receive. Here is how to read it, and how to use it to turn the programme around.

11 min read

Read →

AI programmes · Delivery

AI Programmes Don't Fail at the Model

80% of AI projects fail. IBM Watson burned $62 million without treating a single patient. Zillow lost $304 million. None of those programmes failed because the technology stopped working. Here is where they actually failed, why the AI industry keeps pretending otherwise, and what it costs when a programme believes it.

11 min read

Read →

Security · AI Building

OWASP Top 10 for Vibe Coders

The average breach costs $4.44 million. Your AI writes the code you asked for, not the security you forgot to ask for. I ran OWASP ZAP on my own production app and found fourteen medium findings and three highs. None were things Claude warned me about.

16 min read

Read →

Building · AI

25 Things I Learned Building Production Apps as a Vibe Coder

From async job queues to feature flags, idempotency to gradual deployment. The lessons that don't show up in tutorials, learned by shipping two production systems with AI assistance.

19 min read

Read →

AI Building · Product

The Spec Was Never Updated

I planned eight AI agents, a seven-stage pipeline, automated critique loops, and differential CV rendering. I shipped two LLM calls and a human checkpoint. The pipeline ran end-to-end for the first time only after fixing eleven production bugs in a single session. The PRD still describes the original architecture. The code disagrees.

15 min read

Read →

AI Building · Quality

The Tests That Passed Without Running

Six Playwright spec files written, committed, and declared done. Five of six had wrong route paths, transcribed from PRD acceptance criteria rather than from the actual code. All six skipped cleanly. The tests passed. They had never run. This is a story about the difference between tests that pass and tests that ran.

13 min read

Read →