notes_on_using_agents_for_everyone/RONACHER.md

Notes from Armin Ronacher Using Claude Code

Sources

YouTubes:

• Claude Vibe Codes a Sentry Clone
• Claude Code Fixes Two MiniJinja Issues
• Agentic Coding： The Future of Software Development with Agents

Tips

• use claude --dangerously-skip-permissions for YOLO mode
• use the playright MCP so Claude can use a browser and make screenshots to see what's happening
• "before implementing this, please think through the problem and come up with a plan"
• for existing codebases, good and fast tools, and observability are critical (see section)
• use O3 to generate architecture first
• you can use other LLM tools via shell tools too e.g. the gemini CLI
  • use this for especially tricky tasks
• (not from Ronacher) you can use /init at any time, it will update CLAUDE.md if it exists
  • #memorize might not work as well (anecdotally)

Conserve Context + Provide Tools + Observability

Conserve Context

• don't use /compact (it's too non-deterministic)
• provide summaries of documentation for the agent (of libraries etc)
• provide a tool in each code base that can make/refresh a summary of all code
• provide a tool to Claude so he can run the above tool and glimpse the summary
• use shell tools rather than MCPs (except Playwright) - context is too short
• provide a tool to look at the last 20 lines of the logs, as well as additional lines when needed

Use subtasks/sub-agents

Provide Tools

• put them all in Makefile
• make it clear how to use a tool, so you don't get
  • silent failures ("all tests ran" even though it found zero)
• early on ask Claude to use a tool and see if it's working well
• for tricky concurrency bugs, give a tool it can put in the codebase to generate logs for debugging

Observability

• use a unified log (a single file with all logs)
  • SQL logs maybe (unless too verbose; otherwise put it in a separate log that's available to CC)
• forward console.log in the browser to a log file
  • https://github.com/mitsuhiko/vite-console-forward-plugin
• make tail-logs

TODO

• Determine what he puts in his systemwide CLAUDE.md
  • https://github.com/mitsuhiko/agent-prompts
  • "here's how you write throwaway code: "

        * When analyzing large codebases or multiple files that might exceed
         context limits, use the Gemini CLI with its massive context window.
        * Use gemini -p when:
        - Analyzing entire codebases or large directories
        - Comparing multiple large files
        - Need to understand project-wide patterns or architecture
        - Checking for the presence of certain coding patterns across the entire codebase
    
        Examples:
    
        gemini -p "@src/main.py Explain this file's purpose and structure"
        gemini -p "@src/ Summarize the architecture of this codebase"
        gemini -p "@src/ Are there any React hooks that handle WebSocket connections? List them with file paths"
    

• get set up with his "unified log"
  • https://github.com/mitsuhiko/vite-console-forward-plugin
• learn how to use .claude/commands
  • ronacher
  • disler
  • ...

small

What are these?

• git sync
• git autotrack

Nuggets

• you can use CC to do web research via playright, for non-coding things too
• you can also use it to do things on the web like selling stuff IRL
• uvx to not have to deal with installing dependencies
• use golang or PHP over Python
• don't commit to an annual subscription of anything because these tools are evolving so much

With AI, every nation, every corporation will want its own models, its own control, its own stake in the future. — https://lucumr.pocoo.org/2025/6/4/changes/

• this is very interesting; it points to where the juice is going to be: custom models, or at least custom agents.