See how the system actually works.

Featured system

Persistent Context & Memory for Coding Agents

Claude Code forgets everything between sessions, so every new chat starts with you re-explaining your project. This architecture builds the fix: an Obsidian vault holding the agent's memory, commands and skills as plain markdown, symlinked back so the agent reads them as its own.

Systems library

Three tracks, one standard: it has to run.

Each track covers a different layer of the same stack. Here is what each one will walk through.

• 01

  AI Workflows

  How an agent setup is arranged so it holds context across a real project — where instructions live, what the model is allowed to see, and how the whole thing is recovered after a restart.

  • context layout
  • instruction files
  • review loops
• 02

  Developer Tooling

  The small tools that remove friction: command-line utilities, editor integrations, MCP servers, and scripts. Built on screen, with the reasoning behind each design decision left in.

  • cli design
  • mcp servers
  • editor integration
• 03

  Automation Systems

  Pipelines meant to run unattended — capture, processing, publishing, monitoring. Wired end to end, including the error handling that decides whether they are safe to leave alone.

  • scheduling
  • failure handling
  • observability

Method

Demonstration over commentary.

Predictions about AI are cheap. A system either runs or it does not, so every breakdown follows the same three steps and ends with something that works.

1. 01

   Problem

   Start from friction that costs real time in real work — a step repeated too often, a handoff that keeps breaking, a task that never quite finishes. The problem is stated concretely before anything gets built.

2. 02

   Working System

   Build it on screen until it runs. Config, code, and wiring in full, with the dead ends and the errors left in, because the recovery is usually the part worth watching.

3. 03

   Reusable Components

   Separate the parts that transfer from the parts that were specific to the example — the script, the config, the prompt, the structure — so they can be lifted into a different setup.