I Ran the Memory Prompt on This Site's Repo. Here's What It Wrote.
Your AI agent starts every session knowing nothing about your project. I've already published the full internals of my memory network . Fair warning applies to that post: it's months of accretion across every project I run, and nobody should start there. So this post is the other end of it. The five-minute version. One paste, and what it actually produced on a real repo. The demo repo is the one…
A recent AI experiment involved running a memory prompt on the repository for the website andrewdetwiler.com. The demonstration started with a fresh agent session having no prior knowledge of the site. After the prompt was provided, the agent generated five small markdown files, totaling about ninety lines.
The main file, CLAUDE.md, outlined what the site was, the real commands, and the important rules: the em dash gate, the everything-is-ungated rule, and the minimal JavaScript policy. Another two-line file, AGENTS.md, served as a pointer to the main file for any tools to locate it.
The agent also created docs/now.md, a document detailing what was currently active by referencing recent commits in the repository. A decisions file, docs/decisions.md, was populated with four decisions that the agent extracted from the history, including the reasoning behind keeping the embargo word list out of git due to the risk of leaking names.
Additionally, the agent produced docs/notes.md, which included three cautionary notes, such as the fact that the pre-commit hook does not function until the git config core.hooksPath .githooks is set. This particular detail was particularly noteworthy, as the agent had discovered it independently without being explicitly instructed.
The agent adhered to the scope of the prompt and did not modify any existing files, instead adding context around the existing contents without replacing them. It recognized the pre-existing context files, such as a glossary and an open-items list, and referenced them from the main file. However, the agent did not generate deep insights; instead, it provided orientation and a foundational understanding of the site's structure and behavior.
The prompt was designed to be ungated and free, allowing others to access it from buildwithamemory.com. The ninety lines of code produced were designed to be plain markdown files within the user's repository, alongside the code they describe. The agent's role is to follow instructions, automatically logging decisions and observations made during the development process. The prompt and demonstration are original works, with the method drawing from various sources recorded in the CREDITS.md file.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
