Show HN: ADRs for AI – Deliberation records for human–AI work, versioned in Git

ADRs4AI is a version-controlled method for preserving the reasoning behind human-AI decisions, so work can resume without relying on anyone’s memory of past conversations. A person starts with a rough “seed” of their thinking; an AI organizes it into decision records that retain the person’s words, alternatives considered, rationale, open questions, and the AI’s recommendation, confidence, and conditions that could disprove it. The human then accepts, revises, or overrides the recommendation. A sample record weighs three policies for correcting historical claims and recommends a bitemporal approach, distinguishing when something was valid from when it was observed; the founder supplies the ruling from a phone. The record attributes each contribution. The method is explicitly not transcript storage, a vector-database memory system, or a way to give AI persistent identity: it preserves the work, not a mind.

The project provides a GitHub starter template, a VS Code extension, and an iOS client. The template includes record formats and workflow guidance; the extension lists open questions, flags malformed records, and lets agents query the records through an MCP server. The phone app supports reviewing questions and accepting answers that become commits. A basic workflow can begin in about ten minutes: create a repository, add initial context, commit before building, give an AI a mission, capture a seed, and answer the questions it generates. Questions are searchable with grep, while coordinating multiple AI collaborators is an optional extension. The central claim is that keeping decisions and their reasons makes collaboration resumable and lets its reasoning accumulate across people, agents, and sessions.