Keep human-AI design decisions as versioned Git records
At a glance
ADRs4AI turns project context, open questions, and accepted or overridden agent recommendations into records a team can revisit. This workflow is listed under software architecture, decision records, human-AI collaboration, Git.
- Agents / tools
- ADRs4AI template, GitHub, Git, Claude Code or another coding agent
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Start a private project repository from the ADRs4AI human-AI collaboration template, fill in the mission block in CLAUDE.md, and commit the initial project record before coding. 2) Ask the coding agent to read the mission and state its understanding or push back before implementation. 3) Dictate a rough idea or seed and have the agent break it into structured decision records with options, a recommendation, rationale, confidence, falsifiers, and open questions. 4) Answer, accept, or override the open questions and commit the decisions; use `grep -rn '^### QST' docs/adr/` to find unresolved questions when returning to the project.
Worth knowing
This is the project's own workflow, presented by its creator. HN commenters raised the risk that records can become stale or conflict with later work, so review and revise them as the project changes.
