How to Build a Self-Improving Claude System

TL;DR

Build a self-improving Claude system by creating a structured knowledge base, ingesting existing personal data, and maintaining tested pipelines for new information. Use reusable skills to process raw resources consistently, update a navigable wiki, and capture insights from Claude sessions, communications, curated content, and periodic reflections without over-engineering the setup.

Transcript

I've been obsessed with building my own self-improving system with clawed code. And after studying Andrea Cararpathy, the anthropic team, and running my own system, I've identified a five-step build framework that lets anyone create a self-improving system with clawed code. Today, I'm walking through all five steps, exactly how to implement them, a... Read More

Key Insights

• A self-improving Claude system is built on a persistent project that stores data and reusable procedures. Its foundation has two parts: a knowledge base for accumulating context and skills that prescribe consistent handling of repeated tasks.
• An LLM knowledge base uses a raw folder for original resources and a wiki folder that references those resources. The wiki functions like a table of contents, helping Claude locate relevant information without reading every stored file for every task.
• A project instruction file is a persistent reminder of how the knowledge framework operates. Documenting the folder structure and expected workflow there helps Claude follow the same organizational rules whenever it works inside the project.
• A reusable skill is a defined procedure for a task performed repeatedly. The add-new-resource skill, for example, places a source in the raw folder, analyzes it, and creates or updates wiki entries that reference the newly ingested material.
• Historical Claude conversations are valuable training data because they contain the user's actual questions, tasks, and working patterns. Analyzing saved session history can reveal lessons and suggest practical skills that reflect demonstrated needs instead of hypothetical workflows.
• Personal ecosystem data can include computer files, email history, meeting transcripts, Slack messages, and published YouTube transcripts. The system can analyze these sources for context, writing patterns, completed work, and potential opportunities to use AI more effectively.
• Skill-driven data ingestion depends on testing each pipeline before automating it. A processing skill should reliably handle raw data as intended, because later improvement loops depend on those skills to produce consistent information without continual manual correction.
• Selective ingestion produces a higher-signal knowledge base than collecting everything available. Curated newsletters, books, blogs, videos, and personal reflections should be chosen for relevance, while raw content should still be captured when an automated connection is not yet available.

Questions & Answers

Q: How do you build a self-improving system with Claude?

Start by creating a persistent project with a raw-resource folder, a wiki that points to relevant raw files, and a project instruction file that explains the structure. Add reusable skills for repeated operations, then ingest historical data and establish recurring pipelines for new Claude sessions, ecosystem activity, curated resources, and personal reflections. Test every processing skill before using it in an automated improvement loop.

Q: What is an LLM knowledge base for Claude?

An LLM knowledge base is a structured project that separates original resources from navigational information. Raw materials, such as transcripts, are stored in a raw folder. A wiki folder contains entries that reference those materials and directs Claude to the right sources. This arrangement works like a book's table of contents, reducing the need to inspect every file whenever information is needed.

Q: Why should repeated Claude tasks become skills?

Repeated tasks should become skills because a skill records exactly how Claude should perform a procedure each time. This reduces back-and-forth and makes processing more consistent. A basic add-new-resource skill can accept a raw file, place it in the appropriate folder, analyze its contents, and create or update relevant wiki entries. Multiple utility skills can later support larger orchestration workflows.

Q: What data should be uploaded to a Claude knowledge project?

Useful initial data can come from three broad areas: prior AI interactions, a personal digital ecosystem, and personal or project context. Examples in the source include Claude session history, relevant computer files, exported email, meeting transcripts, Slack messages, YouTube transcripts, a recorded life story, and stated project goals. Privacy-sensitive material does not need to be included if the user is uncomfortable sharing it.

Q: How can Claude session history improve an AI system?

Claude session history records the user's real questions, recurring tasks, and ways of working with AI. Analyzing that history can produce clear lessons and suggestions for new skills, especially when the analysis explicitly asks how the system could improve. A recurring synchronization skill can continue bringing newer sessions into the project and processing them, rather than treating conversation history as a one-time upload.

Q: What recurring data pipelines should a Claude system use?

The framework identifies four pipeline categories: the user's own Claude inputs, recurring personal ecosystem data, curated external content, and periodic data dumps. These can be handled by dedicated skills that pull new material and process it into the project. Examples include session synchronization, meeting and Slack capture, newsletter extraction, public YouTube transcripts, and spoken summaries of lessons learned during a day or week.

Q: How should curated content be added to a Claude knowledge base?

Curated content should come from useful external resources such as books, blogs, YouTube videos, or topic-specific newsletters. A synchronization skill can pull new material, extract key claims, and process those claims into the wiki. Email aliases can help filter newsletters into a distinct location. The source recommends being selective because the goal is high-signal context, not the largest possible collection of information.

Q: Why must ingestion skills be tested before automation?

Ingestion skills must be tested because automated improvement loops depend on them to process raw information correctly and consistently. If the underlying skill does not work as intended on the user's machine, automation can repeatedly introduce poorly organized or incorrectly processed material. The recommended sequence is to define the skill, verify its behavior with actual data, and only then use it as part of a recurring pipeline or broader orchestration process.

Summary & Key Takeaways

• The BASE step creates the system's foundation through a knowledge base and reusable skills. Raw resources go into a raw folder, while a wiki points Claude toward relevant files like a table of contents. A project instruction file documents this structure, and an add-new-resource skill standardizes future ingestion and wiki updates.

• The UPLOAD step turns historical information into useful project context. Recommended sources include past Claude sessions, relevant computer files, exported email, personal history, and project goals. Skills should emerge from observed work patterns rather than hypothetical needs. Privacy-sensitive sources can be excluded, while recorded interviews can fill gaps in personal context.

• The INFLOW step keeps the knowledge base current through recurring pipelines. Separate skills can synchronize Claude sessions, ecosystem activity, curated external content, and spoken reflections. Each pipeline should preserve raw information, process new material consistently, and update the project. Selectivity matters because a smaller collection of high-signal resources is more useful than indiscriminate ingestion.