There is a peculiar pattern when working with Claude Code. It is perfect for the first 30 minutes, but after about an hour, it forgets the agreed-upon patterns and starts building components again. This phenomenon is called context rot. GSD solved this problem by launching a fresh, clean AI for every task and externalizing the context into a file. The Max plan is recommended due to high token consumption, and it is overhead for simple tasks. Let's go through the entire cycle using a to-do list app as an example.
You must have Claude Code installed first (https://claude.ai/code). Node.js 18 or higher is also required. Please check first if the claude command works in the terminal.
To introduce GSD into an existing codebase, run /gsd:map-codebase before /gsd:new-project. Parallel agents analyze the existing stack, architecture, and conventions to help avoid conflicts with existing patterns.
For questions you are unsure about, you can say "figure it out yourself." GSD will select reasonable default values. If you add the --auto flag, GSD reads the codebase, makes its own judgment, and only asks about unusual questions.
Please make sure to open PLAN.md directly and check it. If you are not satisfied with the plan, you can request modifications at this stage. It is much cheaper to change it here than after execution.
If an error occurs during execution, GSD will automatically retry. If that still doesn't work, using /gsd:debug (built-in command) will allow the debug agent to identify the cause. Do not close the terminal while running.
Simple modifications can be processed quickly without the full cycle using /gsd:quick 'description'. Pause the task with /gsd:pause-work and resume it with /gsd:resume-work. To save tokens, you can lower the sub-agent model with /gsd:set-profile budget. If the legacy code is complex, you may need to manually fix the .planning/ documentation.
There is a peculiar pattern when working with Claude Code. It is perfect for the first 30 minutes, but after about an hour, it forgets the agreed-upon patterns and starts building components again. This phenomenon is called context rot. GSD solved this problem by launching a fresh, clean AI for every task and externalizing the context into a file. The Max plan is recommended due to high token consumption, and it is overhead for simple tasks. Let's go through the entire cycle using a to-do list app as an example.
You must have Claude Code installed first (https://claude.ai/code). Node.js 18 or higher is also required. Please check first if the claude command works in the terminal.
To introduce GSD into an existing codebase, run /gsd:map-codebase before /gsd:new-project. Parallel agents analyze the existing stack, architecture, and conventions to help avoid conflicts with existing patterns.
For questions you are unsure about, you can say "figure it out yourself." GSD will select reasonable default values. If you add the --auto flag, GSD reads the codebase, makes its own judgment, and only asks about unusual questions.
Please make sure to open PLAN.md directly and check it. If you are not satisfied with the plan, you can request modifications at this stage. It is much cheaper to change it here than after execution.
If an error occurs during execution, GSD will automatically retry. If that still doesn't work, using /gsd:debug (built-in command) will allow the debug agent to identify the cause. Do not close the terminal while running.
Simple modifications can be processed quickly without the full cycle using /gsd:quick 'description'. Pause the task with /gsd:pause-work and resume it with /gsd:resume-work. To save tokens, you can lower the sub-agent model with /gsd:set-profile budget. If the legacy code is complex, you may need to manually fix the .planning/ documentation.
3 comments
You need to log in to write a comment.