Zero manual code: spec-first development in action with Claude Code
A developer built an open-source agent named Agent Smith and a spec-first development plugin to generate 500k lines of code and 10,000 tests entirely via AI with zero manual coding. This approach achieves multi-agent parallel development using YAML specs, isolated Git worktrees, and strict quality cards.
Reddit r/ClaudeCode/u/holgerleichsenring
Claude CodeAgent Smith
Added 2026-10-31 · originally published 2026-10-27
Steps
- 1Create a YAML spec file with a fixed schema in the repository, including goals, steps, and a definition of done, with states cycling among planned, active, and done.
- 2Write decision records to a separate decision file so subsequent agents can read previously built content and the reasoning behind it.
- 3Before writing code, use a script to check whether all files and lines referenced in the spec exist, then have a second agent review the spec to yield a review result of BUILD, BUILD WITH CUTS, or REFUSE.
- 4Assign an independent git worktree to each spec so multiple agents can run in parallel overnight without interfering with one another.
- 5Configure a commit hook to run the full test suite on every spec commit, blocking the commit if any tests fail.
Copy-ready prompt
Based on the following YAML spec file [YAML file path] and decision file [decision file path], accomplish the goal while following the definition of done. First verify via script that all referenced files and lines exist, then begin writing code and ensure all tests pass.
#AI Coding#Agent Practice#Automated Development#Claude
Related use cases
Summarized by AI from the original source - refer to the original post for details. Copyright belongs to the original author.