Turn every correction into a memory your AI keeps
Ready nowfeedback_*.md files written after every correction, a MEMORY.md index under 24 KB, canon files per workspace, and the weekly-review, self-improve and fix-brain skills that propagate lessons across workspaces.
Kit
workspace-os
You need
| Account | Why | Free or paid | Plan |
|---|---|---|---|
| Claude Code (Anthropic Claude Pro or Max) | The AI operator every system runs through. Skills, hooks, memory, MCP servers and the Brain plugin all load inside Claude Code. | Paid | Claude Pro (20 USD/mo) works for light use; Max (100 USD or 200 USD/mo) for daily multi-hour sessions. Anthropic terms require each user to hold their own subscription; never share a login. |
What you bring
- Your willingness to correct the AI in chat, which is itself the input
The steps
- Find the memory folder
- Write the memory rule into CLAUDE.md
- Write the first feedback file from a real correction
- Test that it survives a new session
- Create the weekly review skill
- Create the fix-brain skill (monthly, or when something feels off)
- Install the reviewer pass prompt
- Escalate repeat failures from prose to machinery
Goes further if you have already built Set up your Claude Code workspace operating system. You can build this without it.
Get the full playbook
Part of: And more