Blog.
Notes from building: architecture, agents, tests and interface details, each one tied to something I shipped.
Artificial Intelligence7 min read
Can Kimi K3.1 Pro replace Claude Opus 5.5 in your coding agent?
Yes for test-checked agent runs: Kimi K3.1 Pro ties Opus 5.5 on Terminal-Bench and costs about a third per task. Opus keeps strict specs.Artificial Intelligence8 min read
When should you clear your coding agent's context?
Clear it on a new task, after two failed corrections, or when it circles. Fresh contexts beat one long agent run by 66% in an October study.Engineering7 min read
Does PHP2Shell put your business website at risk?
Yes, if it runs on PHP and nobody updates it: a third of PHP sites run versions PHP no longer fixes, and hacked sites sell casino links.Product8 min read
How to scope an MVP so the first version ships
How to scope an MVP: cut what needs something new, note where it plugs in. Two agents cost Relay 11 existing files; ElevenLabs, 40.Artificial Intelligence9 min read
Can MiniMax M3.1 Flash fix a real bug for free?
MiniMax M3.1 Flash fixed a real bug in 3 of 3 runs for $0, in about 4 minutes each. DeepSeek V4.1 Flash took 27 s for 1.4¢.Engineering10 min read
How to move to a modular monolith without a rewrite
A tangled codebase can become versioned modules without a rewrite: cut import cycles in place first. How SharpOS got to 53.Artificial Intelligence8 min read
DeepSeek V4.1 Flash test: 3 of 3 runs green, under 2 cents
Can DeepSeek V4.1 Flash fix a real bug? 3 of 3 runs passed 18 failing time-zone tests, in about 90 s and under 2 cents each.Artificial Intelligence11 min read
Too many MCP tools? 130 tools for half the context tokens
Too many MCP tools? On SharpOS, 130 tools cost about 32,000 tokens up front, half a strict mirror, with one registry and lazy schemas.Engineering10 min read
TDD with coding agents only works behind a required gate
A developer directing coding agents can keep a large codebase correct with a failing test as the spec and one required gate on main.