Blog
- 016: Vision Landing
Rewriting the landing page as a manifesto — from feature list to product vision.
- 015: Vertical Iterations
Why every iteration cuts through the entire stack — code, docs, blog, tooling — instead of building one layer at a time.
- 014: Teaching the AI Your Codebase
Splitting a monolithic CLAUDE.md into per-repo files so the AI assistant has the right context wherever it works.
- 013: Visual Identity
One color palette across the landing page, docs, and blog. The landing page gets simpler, the docs stop looking like a default template.
- 012: Agent Gets a Web Server
Adding nginx to every agent container — the base layer owns the web server, adapters extend it through conf.d, and nginx specificity rules make the whole thing composable.
- 011: The Blog Is Infrastructure
The publishing system behind Reflection's engineering blog — Astro Content Collections, structured metadata, a visual style guide, and a process that makes publishing part of every iteration.
- 010: Vision & Roadmap
Nine iterations of infrastructure crystallize into a product vision: a platform for founding and managing AI-native companies.
- 009: Second Adapter
One adapter is an implementation. Two adapters is a pattern. ZeroClaw proves the capsule abstraction holds across completely different runtimes.
- 008: Persistent Home
Docker named volumes, a self-healing entrypoint, memory limits, and graceful shutdown. The agent's state survives redeploys.
- 007: Readable Repos
Seven repos, seven READMEs. Consistent format, no boilerplate. The project goes from 'read the Nix' to 'read the README' in one iteration.
- 006: Demo Deploy
A dev launcher watches the capsule's git repo, builds in an isolated git worktree, and restarts the container automatically. Push a change, the agent updates itself — safely.
- 005: Agent Remembers
The agent gets session persistence. Each Telegram chat maps to a UUID, two-phase invocation handles failures gracefully, and ~40 lines of bash give the agent conversational continuity.
- 004: Agent Speaks
The agent gets a Telegram transport. Same config, different adapter import. One line changes, silent container becomes a chatbot.
- 003: Config to Container
nix build turns a capsule into a Docker image. Non-root user, writable home, zero Dockerfiles.
- 002: Docs From Day One
Most projects add docs when they're ready. That's backwards. A standalone Starlight site deployed at iteration two.
- 001: The Agent Is a Repo
An agent defined in 17 lines of Nix. A schema that validates identity at eval time. Two repos, zero frameworks.
- 000: What if an AI Agent Was Just a Git Repo?
Every AI agent framework wants you to write code their way. What if the agent's identity just lived in a Git repo instead?