AI Second Brain #ai #natebjones #building #blocks

ai_second_brain.md

https://www.youtube.com/watch?v=0TpON5T-Sw4

Second Brain

• Not designed to be a storage system
• Relationships that cool off
• Hold 4-7 things in memory = Bad at pattern recognition where patterns are scattered
• AI loop continues whether you are motivated today
• Non-engineer: Slack + Notion + Zapier + Claude/ChatGPT
  • Databases
  • People
  • Projects
  • Ideas
  • Admin
  • Inbox Log

Building Blocks

- Drop Box/Capture Door/Ingress Source
	- Slack Channel
- Sorter/Classifier/Router
	- Taxonomy at execution time 
	- Classification is a solved problem
- Form/Schema/Data Contract
	- Kind of data you store for each kind of thing
	- Makes automation possible
- Filing Cabinet/Memory Store/Source of Truth
	- Writable by automation
	- Readable by humans
	- Supports simple filters and views
- Receipt/Audit Trail/Record
	- Trust
	- Errors feel mysterious
	- Inbox Log database
	- Tracing
- Bouncer/Confidence Filter/Guardrail
	- Prevents low quality outputs from polluting memory store
	- When Claude/ChatGPT classifies your thought it also returns confidence score (0-1)
		- Below threshold classified as "Needs Review"
	- Keeps Second Brain from becoming a junk drawer
- Tap On The Shoulder/Proactive Surfacing/Notification
	- Nudges
	- Makes system alive
	- Puts the right info in your path so you don't need to look for it
	- Sunday Review/Weekly summary from Zapier
- Fix Button/Feedback Loop/HITL Correction
	- One-step way to correct mistakes without opening dashboards
	- Corrections must be trivial or ppl will not make them
	- 

Building Principles:

- Reduce the human's job to one reliable behavior
	- Scalable move is to make the human do one thing
	- Human's job is to throw thoughts at the system and that's it
- Separate Memory From Compute and Interface
	- Most important architectural principal
	- Memory is where truth lives (Notion)
	- Compute is where the logic runs (Zapier/Claude)
	- Interface is where the human interacts
	- Makes the system portable and swappable
- Treat Prompts Like APIs not Creative Writing
	- Scalable agentic prompt is a contract
	- Fixed input format (Schema)
	- Fixed output format (JSON)
	- We don't want model to be helpful in uncontrolled ways
	- Reliable beats creative in these systems
- Always Build a Trust Mechanism
	- Not just a capability
	- Confidence scores
	- Fix button to correct things in trivial manner
- Default to Safe Behavior When Uncertain
	- Real agentic system must know how to fail gracefully
	- Safest
		- Just hold, don't file
		- Log it
		- Send email to ask for clarification
- Make Output Small, Frequent, Actionable
	- Want terse list
	- Reduce cognitive load
	- Outputs should be small and on set cadence
- Use Next Action As Unit Of Execution
	- Most project notes fail because they store intentions not actions
	- "Work on website" is not executable action
	- "Email Sarah to confirm the project deadline" is executable action
	- Database needs field called "NextAction"
	- Classification prompt needs to be tuned to extract very specific actions
- Prefer Routing Over Organizing
	- Humans hate organizing
	- Humans like to drop things in a box and forget about them
	- Let the system (Claude) route to a set of stable buckets
	- Go for four buckets
- Keep Number of Categories and Fields Painfully Small
	- Richness causes friction
	- Friction kills adoption
	- Start with minimal fields
	- Add sophistication later
	- Start simple and stay simple
	- Only add complexity when evidence says its needed
- Build System For Restart Not Perfection
	- Scalable system should be easy to restart
	- Don't catch up: restart instead
- Build One Workflow And Then Attach Modules
	- Add optional capabilities later
	- Minimum viable loop
- Optimize For Maintainability Over Cleverness
	- Moving parts are failure parts
	- Optimize for 
		- Fewer tools
		- Fewer steps
		- Clear logs
		- Easy reconnects
			- Zapier token expires and stops workflow you want to restart and move on
			- Not debug for an hour