A new Git tool records human–AI decisions so you don't lose the reasoning behind them
ADRs4AI lets teams version-control their back-and-forth with AI, so every choice keeps its 'why' for later.
What you could do with it
- A developer or knowledge worker could capture their back-and-forth conversations with an AI and turn them into permanent Git-tracked decision records so the reasoning is never lost when the chat window closes 1.
- A project founder could review open questions generated by an AI on their phone and tap one Accept to log their ruling as a Git commit without needing to be an expert engineer 1.
- A research team could keep a living, version-controlled record of every design choice and its rationale so future collaborators can understand why decisions were made and what would prove them wrong 1.
Your AI conversations now leave a paper trail
When you work with an AI, the thinking it produces — options it considers, recommendations it gives, the confidence behind each — normally disappears when the chat window closes. ADRs4AI stops that from happening by turning every session into a file stored in a Git repository, the same system developers use to track changes in code 1.
Each record follows a simple format. It states the question, lists the options the AI considered, includes the AI's recommendation, its rationale, how confident it is, and even what would prove the recommendation wrong 1. Then a human reviews the record and accepts, overrides or edits the answer . The result is a living document where both sides of the conversation are preserved.
One messy note becomes a structured record
The workflow starts with a seed — just type or dictate your thoughts, however messy 1. The AI then organizes that dump into ADRs that keep your actual words along with the options, reasoning and open questions . Every question is tagged with a prefix like "### QST:" so it can be found instantly by searching, or surfaced automatically in the VS Code extension .
A real example comes from the frontier-model-observatory project, where an ADR asked how corrections should preserve what was previously believed 1. GPT 5.6 Sol, running through Perplexity Computer, produced the options, rationale, confidence score and a falsifier — the test that would prove the recommendation wrong . The final one-line ruling came from founder Jérémie Lumbroso on his phone via the ADRs4AI iOS app . That is the whole mechanic in practice: the AI carries the reasoning, the human spends one judgment call, and the record keeps the why .
It takes ten minutes to start
The creators say a first project can be set up in about ten minutes 1. You create a private repository from the ADRs4AI template, fill in a top-level CLAUDE.md file, run a few git commands to commit the initial record and then invite your agent to take on the mission . From there, you dump thoughts into a seed, answer the first question and repeat the loop .
What it is and what it is not
ADRs4AI is not designed to archive every line of a chat, store memories in a vector database, act as an orchestration platform or claim that AI has a persistent identity 1. It simply records what was decided, why and what remains open — so future you, or anyone else on the project, can follow the thread without starting from scratch . The creator says the method was built through daily practice with AI collaborators and is meant to work for beginners as much as for experienced engineers .
This story reports on a trend for information only. It is not advice or encouragement to do any of it.
Key facts
- ADRs4AI stores human-AI decisions as version-controlled files in Git
- A VS Code extension surfaces open questions with diagnostics and deep links
- An iOS app lets users review and answer records on the phone
- Records include options, recommendations, rationale, confidence and falsifiers
- The loop takes about ten minutes from an empty repository to a first answer