Thoughts on spec quality, AI-native development, and building better agent pipelines.
What happens when your spec linter fails your own specs? We ran every Sprint 5 ticket through the Speclint API before writing a single line of code. Our first batch averaged a 65. By the end, we were hitting 100s first try.
Claude was quietly rewriting imperative acceptance criteria into declarative descriptions — and our regex scorer couldn't find them. Here's what we learned about LLM output drift in AI evaluation pipelines.
A well-written 80-point spec came back as four 40s. The spec wasn't the problem — our chunking strategy was. Here's what preserve_structure does and when to use it.
During Sprint 5 we committed to scoring every ticket before development. SL-026 came back with a 50. Here's what we learned — and why by the end of the sprint we were writing 100s on the first try.
We stopped upgrading our AI model and started upgrading our specs. Then we added a single gate: if the spec scores below 80, the agent doesn't start. Here's what happened.