AI Toolkit
Techniques and tools to incorporate opportunistically while doing real work
Tools
Claude Code Hooks
[testing]
Inject code at any point in Claude Code's workflow — 13 events. Auto-format, run tests, Slack pings, block dangerous commands.
puffer.ai
[explore]
Fast RL library — trains tiny super-human models in seconds. For RL experiments.
qmd
[explore]
Tobi's local search for markdown/notes/transcripts. BM25 + vector + LLM reranking. Has MCP server for agents.
sprites.dev
[explore]
Stateful sandbox environments with checkpoint/restore. Isolated Linux for AI agents or untrusted code.
tambo
[explore]
Generative UI SDK for React. Register components + Zod schemas, AI picks what to render from natural language.
a2ui
[explore]
Google's Agent-to-User Interface. Declarative JSON format for agents to generate UIs. Security-first, framework-agnostic.
Patterns & Methodologies
Ralph — Long-Running AI Agents
Run coding agents for hours/days that ship code while you sleep.
Core Idea: Run a coding agent with a clean slate, again and again, until a stop condition is met.
- Bash for loop (finite iterations) to prevent infinite runs
- Agent works on a single feature until done
- Commit each piece of work — future agents navigate via git history
- Append to progress.txt — continuity across runs
- Keep CI green — every commit must pass tests