A spec-first workflow reports 1,332 specs and a commit gate that blocked work before it reached a branch

Leichsenring builds Agent Smith, an open source tool that turns tickets into pull requests, with Claude Code and a rule that he writes none of the code. His report, with figures he counted from the repository on 28 October 2026, lists 1,332 finished specs, 743,663 lines of C# and 10,944 tests. Each change starts as a YAML spec. Its claims about existing code cite a file and line, a script resolves each citation, a fresh agent reviews the spec, and a commit hook runs the full test suite and blocks the commit when anything fails. Since the hook began keeping a ledger on 24 August, he reports 111 of 660 specs were blocked at least once before they could commit.

The failure list is the useful part. Specs he wrote from his own beliefs about the code kept being wrong: across two adversarial reviews in September, fifteen rounds in all, "every round found at least one stated fact that the code contradicted." Markdown specs filled with stale code samples; 93 of them came to 830 KB, and the same 93 as schema-bound YAML came to 171 KB. He also reports that his time spent writing code fell from about 80% last year to 5 to 10%.

Why it matters: For anyone running agents unattended, the portable parts are mechanical: require file and line evidence for claims about existing code, resolve it with a script, and stop the commit rather than the merge. His own advice is "Block the commit, not the merge." These are one builder's figures for one C# codebase, and we did not check the repository counts. They show the workflow holding up at that scale; they do not separate the method's effect from the builder's skill.