◆ Writing
Notes on building.
Honest writing on AI systems engineering. Not thought leadership. Not tutorials. Observations from building production systems: what worked, what failed, and why.
Stress-testing your architecture with AI before you build it
The most expensive bugs are the ones baked into an architecture before a line of code is written. Claude and Gemini are remarkably good at finding them — if you ask the right questions.
What to look for when hiring an AI systems architect
Most companies hire AI architects wrong. They look for model expertise when they should be looking for systems thinking. Here is what actually separates good from great in this role.
My actual AI engineering workflow in 2026 — hour by hour
Not a list of tools. A real account of how I structure a working day using Claude, Gemini, Manus, and other AI tools together. What runs when, why, and what I have stopped doing manually.
The RAG chunking problem that costs most teams months
Most RAG implementations fail for the same reason: naive chunking that destroys document structure. Here is what to do instead, and how to measure whether it worked.
How to measure LLM quality in production (not just at benchmark time)
Benchmark scores tell you what a model can do on a test day. Production metrics tell you what it does to your users every day. These are different measurements of different things.
Agentic development is a discipline, not a tool
Claude Code and similar tools are widely available. The discipline required to use them at production speed is not. Here is what it actually takes.
When AI makes you slower: the tasks to keep manual
The honest answer to where AI assistance has negative ROI. Not every engineering task benefits. Knowing which ones do not is as important as knowing which ones do.
Building a Vanta competitor in 21 days with Claude Code
A precise account of how TraceLayer was built in three weeks using Claude Code as the primary development environment. Not a tutorial. A report from the field.
What to do with a million-token context window
Gemini 2.0 Pro's context window changes which problems are tractable. Here is how to use it well — and the failure modes that waste it.
AI for the parts of engineering nobody talks about
Everyone discusses AI for code generation. Almost nobody discusses it for incident reports, ADRs, onboarding docs, and changelogs. These are where the compounding value lives.
Multi-agent systems are easier than you think. They are rarely the right answer.
The technical implementation of multi-agent LLM systems is well-understood. The decision of when to use them is not.
What Manus is actually good for — a practical guide for engineers
Manus is a web agent, not a coding assistant. Engineers who understand the distinction get dramatically more out of it. Those who do not are disappointed within a week.
Context engineering is what matters now, not prompt engineering
The era of clever prompting is over. The skill that separates engineers who get consistent results from AI is managing what the model knows — not how you ask.
Claude Code as IDE — the 2026 workflow
How I actually use Claude Code day-to-day as my primary development environment. Tools, workflows, habits, and where it breaks.
Writing system prompts and CLAUDE.md files that actually work
Persistent instructions for AI tools are a craft. Most engineers write them once, badly, and never revisit them. Here is how to write system prompts and CLAUDE.md files that compound over time.
Claude, Gemini, Manus: choosing the right model for the right task
Claude, Gemini, Manus, and GPT-4o are not interchangeable. Each has a task profile where it outperforms the others. Here is how I decide.
How to debug with AI faster than you can think
Debugging is the task where AI assistance has the highest ROI and the most underused potential. Here is how to use Claude and Gemini to cut debugging time by an order of magnitude.