NemoClaw Knowledge Wiki

Use Claude code for more than coding eg for research

4 min read

Use Claude code for more than coding eg for research

https://www.youtube.com/watch?v=RV9vBRtXagQ Here is a summary of the video titled “How to Make Claude Code Cook”, detailing how to repurpose Anthropic’s coding tool into a general-purpose autonomous agent system.

Hot Take: Claude Code isn’t just for Coding

The core premise of the video is that Claude Code is essentially an agentic loop wrapped in a coding-focused system message. By changing the objectives and instructions you provide, you can transform it from a coding assistant into a general-purpose automation tool capable of handling research, data analysis, content creation, and file management.

The “Open Agent System” Architecture

The speaker demonstrates a custom file-based architecture called the Open Agent System. Instead of writing complex software, this system relies on Markdown files to define “agents.”

How it Works

  1. **INSTRUCTIONS.md**: The master file (or router). When Claude Code starts, it reads this file to understand what agents are available and what “triggers” (keywords) activate them.
  2. Progressive Disclosure: To save context window space, Claude doesn’t load every agent definition at once. It only reads specific agent files when a trigger word implies that specific agent is needed.
  3. Agent Files: Each agent is a Markdown file defining:
    • Purpose: What the agent does.
    • Triggers: When to use it.
    • Tools: What shell commands or scripts it can run.

The Demo: Video Game History Project

The speaker runs a workflow to generate an interactive HTML history of video games without writing a single line of code manually:

  1. The Researcher Agent: Scours the web for the history of video games based on a source prompt.
  2. The Image Curator Agent:
    • Reads the research.
    • Searches for relevant images.
    • Validates the image URLs using a custom script it wrote to ensure they aren’t broken.
  3. The HTML Generator Agent: Compiles the text and images into a styled, interactive “Arcade Cabinet” HTML card viewer.
  4. The Instagram Agent (Created Live): The speaker asks Claude to create a new agent on the fly. Claude reads the system structure, creates the definition file, creates the slash command, and immediately uses the new agent to generate a social media post about the content.

Practical Example: The “Jingle App”

To prove this works for arbitrary tasks, the speaker updates a personal web application that plays a daily jingle.

  1. The Scenario: He invents a holiday (“National Nothing Day”) and generates a song using Suno (an AI music generator).
  2. The Workflow:
    • He drops the MP3 into an inbox folder.
    • He tells Claude Code to “Ingest Jingles.”
    • The Agent Actions: Claude Code converts the audio (using ffmpeg), creates a visual theme for the app based on the holiday “vibe,” generates a background image, updates the config files, and pushes the changes.
  3. Result: The web app is fully updated with the new content automatically.

Anthropic’s “Agent Skills” vs. Custom System

Anthropic recently released a feature called Agent Skills, which allows similar functionality.

FeatureAnthropic Agent SkillsCustom “Open Agent” System
IntegrationNative, elegant, highly shareable.Manual setup via Markdown files.
CompatibilityClaude Code Only.Universal. Works with Claude Code, Codex, Cursor, Gemini CLI, etc.
Context UsageLoads all skill definitions into context immediately (can be token heavy).Uses Progressive Disclosure (loads instructions only when needed).

The speaker notes that while “Skills” are the future for Claude, the custom Markdown approach is currently more flexible across different AI coding tools.

How to Try It (In < 3 Minutes)

You don’t need to manually build the files to start. The speaker provided a GitHub repository that Claude Code can read to bootstrap itself.

  1. Open Claude Code in your terminal.
  2. Paste the Repo URL: Provide the link to the Open Agent System repository.
  3. Prompt: Tell Claude: “I want to create an agent that can manipulate images.”
  4. Auto-Configuration:
    • Claude Code reads the repo’s OpenAgentDefinition.md.
    • It understands the architecture.
    • It automatically creates the folders, instruction files, and the specific image manipulation agent you asked for.

Conclusion: You can use this method to build agents for meeting notes, data extraction, file organization, or any computer task, essentially treating your CLI as an autonomous employee.   

https://github.com/bladnman/open-agent-system

Graph View