The Straion Blog

Insights on AI-assisted development, coding standards at scale, and building software that stays on track.

Latest off the Press

Developers Asked Anthropic to Build This. We Did.

June 2, 2026

In December 2025, a developer filed GitHub issue #14467 on the anthropics/claude-code repo describing exactly the problem Straion solves. We built it. Here's the full story and why a scattered CLAUDE.md file was never going to be enough.

Katrin Freihofner

Developers Asked Anthropic to Build This. We Did.

Build vs. Buy: The Real Cost of AI Coding Governance

May 19, 2026

Building your own AI coding governance layer sounds like a simple config project. It isn't. Here's an honest breakdown of what it actually takes and when buying is the smarter call.

Katrin Freihofner

Build vs. Buy: The Real Cost of AI Coding Governance

AI Coding Agents

90% Rule Adherence

March 26, 2026

When Claude Code runs with Straion, it stops guessing and starts coding like your best engineer. Straion's automated context layer drives rule-adherence into the 90%+ range, so your AI isn't just writing code, it's following your standards.

Katrin Freihofner

AI Coding Agents

One million tokens won't save your engineering standards

March 19, 2026 · Updated May 21, 2026

The industry is celebrating one million token context windows as the solution to AI coding problems. The research said the opposite in March. Since then, Sourcegraph published 1,281 agent runs across 40+ large codebases confirming it: more context degrades performance. Here's why less context wins, and what the new data changes about how we should build AI coding tools.

Lukas Holzer

One million tokens won't save your engineering standards

AI Coding Agents

Your AGENTS.md is costing more than you think

March 3, 2026

ETH Zurich researchers tested whether AGENTS.md and CLAUDE.md files actually increase the success rate of coding agents. The result: LLM-generated context files reduced resolve rates by ~3% and inflated cost by over 20%. Static context files have no concept of relevance, dumping every rule on the agent regardless of the task. The fix isn't writing a better file; it's changing the delivery mechanism entirely. Deliver only the rules that matter for the specific task, dynamically, at the moment the agent needs them.