Живые скиллы для Claude Code, Gemini и Codex | Как настроить super skills для ИИ агентов
The video explains how to create 'living' or 'super' skills for AI agents like Claude Code, Gemini, and Codex that self-update based on user feedback. The presenter demonstrates how these skills collect feedback after each task, store it in log files, and periodically incorporate that feedback into the skill's core instructions. This approach reduces token consumption and ensures skills continuously improve over time.
Summary
The presenter introduces the concept of 'living skills' or 'super skills' — self-updating prompt files for AI agents that work across Claude Code, Gemini CLI, and Codex. The core problem identified is that traditional AI agent skills are static: they are created once as markdown files, placed in a folder, and never updated, meaning the agent repeats the same suboptimal behavior every time a task is run.
A secondary problem discussed is token consumption. When many skills are loaded into the main Claude.md prompt file, every agent launch consumes large amounts of tokens even for skills rarely used. The presenter's earlier solution was an 'MD register' — a separate file listing skills with brief triggers, so full instructions are only loaded when needed. Living skills extend this by also eliminating dead weight from skills that are outdated or unused.
The living skills system works as follows: after completing a task, the AI agent asks the user for feedback (e.g., which video ideas worked and which didn't). The agent logs this feedback into dedicated feedback.log files stored alongside the skill file. Then, on a scheduled basis (twice a week in the presenter's setup), the agent reads accumulated feedback and automatically rewrites the core skill instruction file to reflect learned preferences. This makes skills 'evergreen' — analogous to evergreen ads that remain effective for years.
The presenter demonstrates a live implementation with two agents: one on Claude Code and one on Codex. He sends both a 9-file documentation package about living skills and asks them to create a YouTube video idea skill. Both agents initially create only a basic skill file without the required feedback infrastructure (feedback.log, additional supporting files), demonstrating that AI agents sometimes need correction. After being prompted to fix this, both agents add the necessary feedback collection files and the system begins working as intended.
The presenter also discusses automating skill execution — scheduling agents to run tasks (like generating video ideas) at night during off-peak hours on Tuesday, Thursday, and Saturday, so reports are ready in the morning. He plans to run this across Claude Code, Codex, and Gemini simultaneously for redundancy and variety.
Key Insights
- The presenter argues that static skills loaded into Claude.md cause rapid token and limit exhaustion on every agent launch, because the agent must process all skill instructions regardless of whether those skills are used in that session.
- The presenter claims that AI systems like Claude Code, Codex, and Gemini CLI are periodically updated with new native capabilities, meaning some older skills become obsolete because the base model can now perform those tasks natively without a skill file.
- The presenter describes living skills as analogous to 'evergreen ads' from his targeting/advertising background — content that never burns out because it continuously adapts to current conditions rather than remaining fixed.
- During the live demo, both Claude Code and Codex initially failed to implement the full living skill system — they created only a basic skill file without the required feedback.log and supporting files, demonstrating that AI agents frequently need correction even when given detailed documentation.
- The presenter uses a cross-agent supervision strategy where, when working with Claude Code, he runs Codex in a background terminal as an overseer to verify whether tasks were completed correctly, and vice versa.
Topics
Full transcript available for MurmurCast members
Sign Up to Access