Writing

Building tools to support your work can quietly become a substitute for the work itself.

Personal AI infrastructure should report results to you, not wait for you to go looking. A small architecture shift changes the whole dynamic.

When you catch yourself setting a reminder to check something later, that's usually a signal that a tool is failing to report what it should.

Adding more AI tools doesn't help if you're still the bus between them.

The difference between a rule that works and a rule that doesn't is usually not the content of the rule — it's where it lives.

Upgrading to a more capable model made my tool sixty times slower. The lesson isn't about models — it's about the difference between capability and fit.

Hong Kong's QDAP annuity is sold on a real tax benefit. But the HK$60K deduction cap is shared with MPF top-ups — and that changes everything.

Andrej Karpathy's autoresearch project is being read as a demo of what H100s can do overnight. It's actually a discipline for doing rigorous work on anything measurable.

Why the structure of an output matters more than the instructions that produce it.

Three hours of debugging revealed two non-obvious behaviours about GPT-5.4-Pro that aren't in the docs: a minimum token budget requirement and a wall-clock timeout gap in Rust async code.

A coding assistant quietly made three productivity apps redundant. Not by replacing them — by making context collapse the boundaries between them.

The resistance to AI agents in banking isn't mostly cultural. It's infrastructure — and the gap is more interesting than the politics.

On why the blank page stopped being the hard part.

WeChat is supposed to be a walled garden. Exa didn't get the memo.

The most sophisticated solution to a problem is usually a sign you haven't found the right abstraction yet.

What changes when LLMs stop being the special thing and become just another software component? The answer is: everything about how you build.

Configuration that belongs to the data shouldn't live in the tool. .gitignore figured this out thirty years ago.

The real ROI of AI coding isn't doing the same work faster. It's doing work that wasn't worth doing before.

The simplest agent architecture might already be the right one: give the agent a file explaining how to do something, and let it read when needed.

Peter Steinberger stopped reviewing AI-generated code entirely. That works for indie software. In regulated environments, it can't. Here's how to think about where you sit.

When behaviour emerges from both code and model responses, reading source files isn't enough. You debug by examining execution traces.

The difference between writing down a rule and making the system enforce it — illustrated by a 15-line hook.

Tool enthusiasm is often net-negative. Context pollution degrades performance faster than features improve it. The principles that actually work.

I kept missing things at the end of AI sessions. So I stopped relying on willpower and systematised the nudge instead.

The instinct to add a rule to a skill file is usually the wrong abstraction. A CLI wrapper enforces at the tool level: zero deliberation, zero token cost.

Most AI governance frameworks are technically-focused risk checklists. Three structural risks are missing from almost all of them.

Most teams build evaluators before reading a single trace. The sequence that actually works is the opposite: observe, categorise, then measure.

OpenAI took the Pentagon contract Anthropic refused. Your AI vendor just became a political statement — and enterprise procurement hasn't caught up.

LLM instruction files are code. They have the same failure modes — with one interesting twist that changes everything.