Summary, pulled from this project's own README: keep a detailed, human-owned record of the actual path from idea to finished work when collaborating with AI — not just the polished result, but the dead ends, backtracks, and resets along the way. Most AI usage today lives inside a company-owned account, where the interaction history belongs to the organization, not the person who did the work. Paul's practice: work through the messy, exploratory part of a project with a personal AI on synthetic data only, then bring the shortest clean path into the actual client or work environment — keeping the full, granular history as his own. Safehouse is the practical, running implementation of that idea: an AI collaborator whose entire interaction history is captured, owned, and inspectable by the person who directed it.
Safehouse on GitHub · Related: OB1
I'm posting because I've spent a lot of time working with AI over the past years and honestly I wanted a place to write this down so I can refer folks when asked what I think. In a nutshell, I believe that from now on, each one of us should build our "AI History" in detail under some secure storage that you own, or at least each of us should plan on collecting and archiving our AI interaction histories somewhere.
AI History is the set of data owned by 3rd party companies, one for each "account" which captures the interaction between a human and an AI model. When working with AI, periodically the model will 'summarize' your session to reduce its memory footprint and provide for a more optimized AI runtime environment. The AI memory is optimized for your specific work. This is how many of us are using AI now.
As I use AI, when I'm working on a project, I will go down many dead ends that didn't work for my goals. When the AI summarizes, you lose the granularity of "how" did you get from inception to product creation. In my humble opinion, keeping that detailed log is critical to understanding the dynamic between the AI model and the human. Today the detailed logs do exist but they are not owned by the human if you are using one of the common models.
As I write, Organizations are using AI in the workplace. Individual contributors are given AI resources such as CoPilot or Claude or Grok or ChatGPT or some home grown AI models. In all cases, the Organization "owns" the accounts/logs and the work product, however, you the Human own the "Human AI Creation Experience" having been the Human side of the effort. In practice, Organizations give AI Agents to Employees to do creative work. All of which is owned by the Organization. However, you, the human are creating with the AI and you are becoming better and better at collaborating with an AI model to do work, you are functioning as an "AI Conductor".
As a result, I have changed how I work in all cases. In practice I am an enterprise scale software architect with mission critical applications running in several sectors. This is what I'm doing for any given assignment:
Any creative work when done by AI/Human "Conductor" will always face the question about insight or what is 'new'. The answer can be discussed in more detail by examining the detailed message history and work products over the creative process, including all the dead ends. Ten years from now, hiring managers may ask to see at least one "Human/AI Composition" from an applicant. A copy of course.
For me this means that 100% of my individual work will be done with me and my personal AIs, never an AI from work. At some point companies will force your hands to use their AIs, but it works in reverse. Use the company AI's but keep your personal AI up to date with concepts you have strung together or created. Never actual data.