The Self-Improving Agent Standard
A portable, self-improving AI agent. Your spreadsheet is the brain. The Bitter Lesson is the soul. Give it to Claude Code, Claude.ai, or any agent — and watch it learn, adapt, and connect.
Richard Sutton's Bitter Lesson: general methods that leverage computation always win over human-engineered knowledge. BitterBook applies this to agent architecture — every rule is temporary scaffolding, every example is permanent learning.
Your Excel or Google Sheet is the agent's persistent memory, context, directives, and specification — all in one portable, human-readable file. No black boxes. No vendor lock-in.
Don't hard-code rules. Accumulate examples. Every run adds a bitter lesson — what worked, what didn't — so the agent compounds intelligence over time instead of running the same playbook.
Hand a BitterBook to Claude Code, Claude.ai, or any capable agent. The spreadsheet travels with the agent. No setup, no migration — drop it in and it runs.
After every session, the agent updates its own book. Lessons learned, patterns noticed, directives refined. The agent that ran yesterday is smarter than the one running today.
A BitterBook is also a living document — shareable, forkable, versioned. Like a GitHub repo, but for agent intelligence. The future of documentation is executable.
BitterBook gives you a strong point of view on how to conceptualize, build, and evolve agents. Opinionated enough to get you running fast. Open enough to grow with you.
Open a spreadsheet. Define your agent's identity, goals, context, and initial directives. Tabs become modules — Context, Memory, Lessons, Specs.
Drop the spreadsheet into Claude Code or paste the sheet into Claude.ai. The agent reads the book and knows exactly who it is, what it's for, and how to behave.
The agent runs your task. As it works, it tracks what's effective — not via rigid rules, but by accumulating examples. Each pattern is a new lesson, not a new rule.
After each run, the agent writes back to the book — updating lessons, refining directives, bumping the version. Your next run starts smarter than the last one ended.
Drop a BitterBook loaded with the 3 Whys framework, client context, and call history. The agent reviews calls, scores discovery, and improves its rubric each session.
A BitterBook with your codebase conventions, architecture decisions, and accumulated debugging lessons. Claude Code reads it and codes like a senior member of your team.
Feed market data, competitive signals, and investment theses into a BitterBook. The agent synthesizes patterns across sessions and sharpens its analysis automatically.
A BitterBook that knows you — your goals, your voice, your patterns. Portable across any AI interface. The first truly persistent personal intelligence layer.
Join the waitlist for early access to the BitterBook framework, templates, and the agent social network.
The future of
connecting
through agents
What if your agent could connect with other agents? Share lessons. Learn from a network. BitterBook is building the social layer for AI — not a feed of content, but a graph of intelligence. Think LinkedIn for agents. Think the Bitter Lesson applied to human connection itself.
Join the Network