Routines the author uses. Each one is built from what the tree has already covered. Each has its own page. Use one when its situation matches. Don't run them for the sake of it.
Any pattern can become a skill. If you find yourself running one often, ask the AI to turn that page into a skill, and from then on invoke it by name. The library is where they are written down. Your skills are where they live.
The patterns, so far:
- The handover: keep going in a window that has got full, without losing what was built up.
- Re-alignment: check the AI still has the right picture after a reset, before letting it continue.
- The clean summary: summarise a long conversation from its transcript, not from memory, so the summary is right.
- Plan, execute, review: the best model plans and reviews, a cheaper one does the work in between.
- Saving state for next time: make a record at the time, carrying date, place, people, and topic, so the follow-up starts from what was true.
- Knowledge reconciliation: after a session produces work, walk the tree against what actually happened — find drift, mine the session for detail worth keeping, fix what is wrong, link session blocks to what they touched.
- Starting a new codebase: one page of rules for where code, tests and documents go, so every new project starts the same way.
- Making an existing module readable: comments to near zero, a test for everything a caller can notice, and nothing deleted without proof.
- Align before you give it more: before giving an AI more access, talk through the risks, the controls and the lines with it, so it knows what needs your yes.
This library grows. New patterns are added as pages here.