Claude Code, better in pair
Improve your AI generated software development plans by running it by one of its peers.
When using Claude Code in terminal, I always have two separate Claude code sessions opened, each with their own context.
When one session gives me an implementation plan, I copy/paste the plan into the other and ask for a review of the plan for any issues.
While the AI attempts to review its own work, during the development of the plan it made assumptions and may have poisoned its own context with those decisions. But when you paste the plan in a new context and ask for analysis, this new context will not have those assumptions known and may point out interesting findings.
As a "Human in the loop", I decide whether the feedback on the plan requires me to go back to the original drawing board, or if I should just provide the feedback to the first session to self-correct.
This is also a really good time to re-evaluate what changes would have been required in the original inputs (your CLAUDE.md, Skills, MCP or your prompt) to avoid the highlighted issues. Sometimes, it means scrapping the entire session, and changing your original prompt to contain more information.
This process is what leads me to use external markdown files most of the time for my tasks, then just ask the AI to read the markdown file to know what is needed. This way, if any corrections are needed, I always have access to the original prompt data (in the markdown file).
I treat AI as a genius brand new developer joining your team. It doesn't matter how good they are, without context they won't be able to succeed. So knowing what to add, or what to remove, is what will make your AI session succeed or fail.
