The Loop That Improves the Loop
Once the baseline loops exist, your attention stops going into doing the work. It goes into improving the system that does the work.
Came across this from @garrytan today, mirrors findings from six months of what I similarly experienced.
I'm doing it for AI systems too however applied across all areas of life which can be codified: from investments to life learnings.
Started with one CLAUDE.md trying to hold everything. Eventually blew out to 10k+ tokens.
The core loop that works:
- A spec skill that governs changes to CLAUDE.md, skills, and agents
- A review agent gating those changes (catches any drift or assumptions being made)
- A synthesizer that owns file structure, naming, and linkages
- A close skill at every session, so the next session picks up cleanly
- Pointers from CLAUDE.md into deeper skills, so the model knows what to reach for
Why this loop works:
Hit something new, the spec skill reads the context and creates exactly what solves it. Output is a new skill. Other skills link into it automatically.
Next iteration of that workflow improves the skill. Fewer steps, fewer corrections, until it becomes repeatable.
The balance: skills hold the decision-making intelligence. Scripts inside them handle the deterministic work. Let AI/LLM handle any decisioning required, whilst anything prescriptive and reproducible is handled by code.
Once the baseline loops exist, your attention stops going into doing the work. It goes into improving the system that does the work. This means you can then focus on things that will bring further value, learning new skills, experiencing new things in life. It becomes a self-perpetuating cycle in itself.
The loop that improves the loop is the real work.