Why I Built TracerKit

In a previous post, I described a workflow for building features with Claude Code: plan in conversation, save the plan as SPEC.md, implement task by task, /clear between sessions. The spec is the memory, the session is the execution. I ended that post saying the workflow would change. It did. Where the Manual Workflow Broke The SPEC.md approach worked, but three problems kept surfacing: Inconsistent specs. Every feature started from scratch. Some specs had decisions sections, some didn't. Some had task checklists, others had prose. Quality depended on how disciplined I was that day. No verification. I'd mark tasks [x] by hand. Sometimes before tests passed. Sometimes I'd forget to update the spec entirely. The source of truth drifted from reality. No lifecycle. Finished specs sat in the project root next to active ones. No archive, no status tracking, no way to see what was in flight. I kept patching these with better prompts and more discipline. Then I realized the workflow itself sho