Claude: The Solution to Avoid Repetition with AI

The Memory Challenge with Claude

Claude, the artificial intelligence developed to assist with various projects, presents a major problem: it does not retain memory of past interactions. Each new session starts as a blank slate, with no recollection of previous discussions or decisions. This lack of continuity forces users to constantly re-explain essential information, a tedious and inefficient process that can quickly become frustrating for those looking to optimize their workflow.

In March, one user decided to switch from ChatGPT to Claude for design projects. While the sessions were generally better, the user quickly noticed having to repeat the same explanations with each new interaction. The projects had been structured using distinct workspaces for each design flow, each with its own context and expectations of Claude. However, this did not solve the problem of repeating personal information and work preferences.

For a while, this user thought that more precise project instructions could resolve the issue. The user tried to refine them, but even with stricter instructions, the need to re-explain the same concepts persisted. The problem was deeper, tied to the very structure of Claude.

The Nature of AI Systems

AI systems like Claude are designed to treat each session independently. They keep no trace of previous exchanges, which means users must continually reintroduce personal and contextual information. This limitation becomes increasingly noticeable as AI is used as a true collaborator, where the quality of the collaboration is crucial.

Faced with this situation, the user began to consider a different solution: what if Claude could remember essential information from each session, instead of starting from scratch? This led to thinking about a method to give Claude a persistent memory, capable of retaining the key information that defines a user's work style and expectations.

The Idea of Identity Files

The solution lies in creating identity files, which provide Claude with a persistent picture of who the user is. These files are read by Claude before each response, regardless of the project at hand. They contain key information about the user's work style, principles, and preferences, allowing the AI to better understand the user and adapt to their needs. This approach aims to turn each session into a seamless continuation of the previous ones, rather than a constant restart.

Components of Identity Files

• about-me.md: This file describes the user's role, current projects, strengths, and what the user expects from Claude. It is a work brief rather than a formal biography, designed to guide the AI in its role as a collaborator. By detailing these aspects, the user enables Claude to understand the context in which they operate and how it can assist them effectively.

• principles.md: This file outlines the user's operating principles, the decisions the user has made, and their approach to trade-offs. It is not a set of directives for Claude, but a frame of reference so that it can reason coherently with the user. By sharing their principles, the user helps Claude align its suggestions and analyses with their way of thinking.

• anti-patterns.md: This file lists common mistakes and failure modes to avoid. It serves as a guide to steer clear of recurring pitfalls in working with Claude. By documenting these anti-patterns, the user reduces the risk of repeating the same errors and improves the quality of interactions.

• voice-samples.md: This file contains examples of passages written by the user that illustrate their communication style. These examples help Claude understand how the user wants their ideas to be expressed. By providing samples of their voice, the user enables Claude to adopt a tone and style that are familiar to them.

How to Create These Files

Creating these files does not require starting from scratch. Users can synthesize existing documents such as case studies, performance reviews, or Slack conversations. The idea is to gather these elements to form a coherent set that Claude can use. This approach makes it possible to capitalize on work already done and to structure the information so that it is easily accessible to the AI.

To do this, users can share this material with Claude during a session and let it interview them section by section. In about 30 to 45 minutes, they can end up with a working system. This collaborative process with Claude makes it possible to refine the identity files to users' specific needs and to ensure that they accurately reflect their work style.

One user has also developed a playbook for setting up the Cowork identity layer. This guide details the entire process, directing Claude to interview users section by section, drawing on their existing writings to build each file. This playbook, along with streamlined versions of that user's four identity files, is available in a GDrive folder.

Automatic System Updates

Once set up, the identity file system updates itself. Claude gradually incorporates the corrections and adjustments made during sessions into the identity files. A weekly scan also detects recurring patterns and integrates them into the system. This continuous updating ensures that the identity files remain relevant and reflect changes in users' work styles.

Over time, this process makes collaboration with Claude increasingly smooth and efficient. Corrections become less frequent, and the AI adapts better to users' expectations. The cumulative effect of these adjustments is a noticeable improvement in the quality of interactions, making each session more productive.

The Benefits of the System

One of the main advantages of this system is that it removes the need to constantly re-explain the same information. Claude gains a sufficient understanding of users' voice and principles to apply this knowledge in new situations without needing to be told again. This saves time and improves the efficiency of work sessions.

Moreover, creating these identity files allows users to better articulate their own ways of working and to identify failure patterns they may not have explicitly recognized before. This process of reflection and documentation benefits not only Claude but also users, helping them clarify and structure their ideas.

In short, this system offers an effective solution for optimizing collaboration with Claude, reducing the time wasted re-explaining essential information and improving the quality of interactions.

Finally, for those who want to deepen their understanding of using Claude as a design collaborator, one user is preparing a companion article. It will explore in detail how that user works with Claude day to day, the persistent project setup, and what a working session with the AI actually looks like. It will be available soon.