Initial notes on working with Claude Code.
Thoughts
- I need better patterns in managing my own context as well as better patterns for managing agent/subagent context.
- The cost of refactor is very cheap now. This can change how our iterative rounds on software products look
- The very begining & end of these AI implemented projects require the closest human attention.
Patterns
- In my user-scoped
CLAUDE.mdI inform the agents to use{project_root}/llm_memoryfor working docs, plans, etc.- This helps prolong working sessions. I can often skip compact when context is low and just point to relevant docs in the folder.
- warning this does require pruning. I have a
/command guiding the orchestration llm to prune, but the trick is remembering to invoke it :)
Workflows
- Achieve longer sessions with
projectssubdir withinllm_memoryfolder