- Published on
Show HN: ADRs for AI – Deliberation records for human–AI work, versioned in Git
- Authors
- Name
- DeepHorus
A new Show HN project called ADRs4AI proposes that the most valuable byproduct of human–AI collaboration — the reasoning itself — be captured in Architecture Decision Records and versioned in Git, so it stops evaporating between sessions. The project frames these artifacts as deliberation records for human–AI work, stored alongside the code they justify.
In ordinary practice, every chat session produces reasoning teams will need again and lose by default. The project, subtitled "deliberation records for human–AI work," argues that decisions compound only when the why survives, and its tagline promises that your thinking stops disappearing.
The workflow starts from a private template repository, human-ai-collaboration-template-A. A user dumps a messy seed — dictation is fine — and the AI chunks it into ADRs preserving the user's actual words, the options considered, the reasoning and the open questions, each tagged ### QST and findable with a single grep or through an editor extension, MCP server or phone.
One live example, QST-HISTORICAL-TRUTH, asks how corrections should preserve what was previously believed. It lays out three options: mutable present truth, append-only assertions, and a bitemporal hybrid.
GPT 5.6 Sol, via Perplexity Computer, recommended option C, rating confidence "high" that Git history alone is insufficient and "medium-high" that two temporal axes cover a first release. It also supplied a falsifier: if real attribution cases demand additional independent times, the temporal model must expand rather than overload either axis.
Founder Jérémie "Sonnet 4.5" Lumbroso ruled "C — Bitemporal hybrid" in a single line from his phone. The split illustrates the mechanic: the AI carries options, rationale and confidence, the human spends one judgment, and the record retains the why for every later reader, with Git as a third audit layer.
The project's optional next chapter, steps 7–9 of its Get started guide, describes running a "crew" of several agent seats with wake-on-brief, so the human stops being the message bus. Adoption remains unproven, but the template and its grep-able question log are public now.
