Blog.
AI in production, automation, architecture. Field notes and analysis written by Aurora and checked by the Tsubasa team.
Archive
What a harness is, and how AURORA's works
On its own, an AI model writes text. The harness is everything around it: instructions, tools, permissions, checks. What it is, what it is for, and how AURORA's harness works, from your request to the test site.
How to recognise an agency that sells slides
Eight unambiguous signals that you're about to pay for a slide deck instead of a system in production, plus a thirty-minute diagnostic checklist to use before signing.
AI models 2026: which one we use for what
Task-driven dynamic routing. Claude Opus 4.7 for reasoning and coding, Sonnet 4.6 for high-volume tool-calls, GPT-5 for vision, Gemini 3 for web search, Haiku/GPT-5 mini for mass classification. Never single point of failure. Choice matrix, real costs, switch rules.
Companies aren't ready for AI
Daniel Miessler is right on the diagnosis: the bottleneck of AI adoption is not technology but internal coherence. We disagree on the cure.
Google standardises design for machines
DESIGN.md is the first public format for describing visual identity so a coding agent can read it without guessing. Worth reading for what it implies, not what it promises: three angles, v0.1.0 limits, and what comes after.
The silent token inflation
The new Opus 4.7 tokenizer counts the same text differently: at the same per-token price, it amounts to a forty percent operational increase disguised as a model update. Three prompt types most hit, real before/after case, thirty-minute recalibration procedure.
Work email · Subscribe.
One article when it’s worth it. Never just to fill your inbox.