# Continual Learning in Claude Code Source: https://www.youtube.com/watch?v=sWbsD-cP4rI Recap page: https://rapidrecap.app/video/sWbsD-cP4rI Generated: 2025-12-30T14:33:45.919+00:00 --- ## Quick Overview The implementation of Continual Learning within Claude Code utilizes 'Skills' as organized folders of instructions, scripts, and resources, allowing the model to automatically discover, load, and improve these skills over time, moving beyond static system prompts by enabling agents to learn from failures and successes through explicit feedback loops like /advise and /retrospective. **Key Points:** - The core problem with current AI agents is that every insight and failure requires manual encoding/intervention, preventing the agent from learning on its own. - Claude 'Skills' are defined as organized folders of instructions, scripts, and resources that Claude automatically discovers and loads, serving as persistent team memory that compounds with every session. - Skills possess key attributes: they are Composable, Portable across Claude apps/Code/API, Efficient (loading only what's needed), and Discoverable automatically. - The learning process involves a 'Learning Loop' with two phases: /advise (before work) to query prior learnings and surface relevant failure patterns, and /retrospective (after work) to auto-generate skills from the session's conversation. - The SKILL.md format requires a 'name', 'description' (which acts as the trigger for discovery), and optional 'allowed-tools' fields. - Anthropic announced official support for Skills in December 2024, including an engineering blog, official skills repository, built-in Claude Code support, and progressive disclosure architecture. - Skills layer and compound, allowing for Personal, Project (repo-specific), and Team (shared via plugins/registry) knowledge to build upon each other. ![Screenshot at 00:03: The slide titled 'The Problem' contrasts the standard five-step manual encoding process for AI agents with the 'Real Cost,' highlighting that the agent never learns on its own, posing the question, 'What if it could?'](https://ss.rapidrecap.app/screens/sWbsD-cP4rI/00-00-03.jpg) **Context:** This video explains the concept of 'Continual Learning' within the context of Claude Code, focusing on Anthropic's 'Agent Skills' feature. This system is presented as a solution to the limitation where AI agents require constant, manual updating of system prompts for every insight or failure they encounter, aiming to create agents that can learn and improve automatically over time, similar to how humans learn from experience. ## Detailed Analysis The video introduces Claude's solution for continual learning via 'Skills,' which are organized folders containing instructions, scripts, and resources that Claude automatically discovers and loads. This contrasts sharply with the current method of developing AI agents, which involves a repetitive, manual process of writing and updating system prompts based on every insight or failure encountered, leading to agents that never learn independently. Skills are designed to be Composable, Portable, Efficient, and Discoverable. The actual mechanism for improvement is the 'Learning Loop,' which consists of two slash commands: '/advise' before work to query the skill registry for past experiments and failure patterns, and '/retrospective' after work to automatically generate a new skill based on the entire conversation, extracting what worked and what failed. The structure of a skill is defined in a SKILL.md file, which requires a unique identifier (name), a description (which serves as the trigger for discovery), and optionally lists allowed tools. The key insight is that knowledge is stored outside the model weights in these skills, allowing them to be read, edited, and shared, and enabling reasoning from one session to compound into future skills. Anthropic announced official support for this feature in December 2024, emphasizing that skills are 'persistent team memory' that compounds every session, rather than being a hack or workaround. ### The Problem - Manual Prompt Encoding vs. Continual Learning: Every insight is manually encoded, every failure requires human intervention, and the agent never learns on its own - The solution lies in enabling the agent to learn on its own. ### What Are Skills? - Organized folders of instructions, scripts, and resources that Claude automatically discovers and loads - Skills are just code (Markdown files) that the model can read and write, meaning the model can improve them automatically every session. ### Skill Structure - Directory layout includes a required SKILL.md file within a skill directory - Storage locations are prioritized: Personal (~/.claude/skills/) > Project (.claude/skills/) > Plugin bundle (Shared). ### SKILL.md Format - Key fields are 'name' (Unique Identifier), 'description' (Triggers discovery), and 'allowed-tools' (Limit capabilities) - The description is the trigger for discovery. ### How Skills Work - Progressive Disclosure involves 1. Claude loads skill names + descriptions, 2. Request matches a skill's description, 3. Asks for confirmation before loading, 4. Full SKILL.md loaded and executes - This scales because metadata is only loaded when needed, requiring no manual invocation. ### The Learning Loop - Before Work uses '/advise' to query skill registry for prior learnings (surfaces experiments, failure patterns, working configurations) - After Work uses '/retrospective' to auto-generate a skill from the session, extracting what worked and failed, and opens a PR to the team registry. ### Stack Your Skills - Skills can be layered and compounded across Personal (workflows/style), Project (repo-specific patterns), and Team (shared via plugins/registry) levels. ![Screenshot at 00:03: Slide detailing the current problem with AI agent encoding: manual steps and the cost of non-learning agents.](https://ss.rapidrecap.app/screens/sWbsD-cP4rI/00-00-03.jpg) ![Screenshot at 00:33: Definition of Skills: Organized folders of instructions, scripts, and resources that Claude automatically discovers and loads, with the key unlock being that the model can improve them automatically.](https://ss.rapidrecap.app/screens/sWbsD-cP4rI/00-00-33.jpg) ![Screenshot at 01:19: Skill Structure layout showing Directory Layout \(my-skill/ with required SKILL.md\) and Storage Locations \(Personal, Project, Shared\) with priority order.](https://ss.rapidrecap.app/screens/sWbsD-cP4rI/00-01-19.jpg) ![Screenshot at 01:55: SKILL.md Format showing key fields: name, description \(the trigger\), and allowed-tools, along with example content for 'What Worked' and 'What Failed'.](https://ss.rapidrecap.app/screens/sWbsD-cP4rI/00-01-55.jpg) ![Screenshot at 02:39: How Skills Work details the 'Progressive Disclosure' workflow steps \(load metadata -\> match description -\> confirm -\> load full skill\) and why this scales \(metadata only when needed, no manual invocation\).](https://ss.rapidrecap.app/screens/sWbsD-cP4rI/00-02-39.jpg)