Skip to main content
File an issue, grab coffee, come back to a PR. Not quite there yet, but close enough to be dangerous. The AI development surface is opinionated, not magical. Each model has a sweet spot. Each tool has a role. Universal configuration lives in one place so every IDE, every agent, every command-line tool reads the same rules.

The tools landscape

Model routing

Each model has a sweet spot, and work is routed to a tier with the needed capability. Deep reasoning and multi-file refactors go to a strong model. Line-level completions go to a fast model, and trivial work can use local inference to limit cloud calls. Model routing covers how that choice is made and governed. One endpoint sits in front of every backend, and callers ask for a role instead of a model name. It also covers the difference between a spend control that is enforced and one that is merely advised. Shared routing rules live in the instructions catalog and arrive through its instructions flake package. See Repo boundaries for how instructions, plugins, and docs are split across repositories.

Repos in this section

Repo boundaries

How rules, plugins, and docs are split across three repositories.

Shared instructions

The shared instruction source used by every AI tool.

claude-code-plugins

Commands, skills, hooks, agents for Claude Code. Project-specific automation.

claude-code-routines

Scheduled remote-agent routines on Claude.ai. Cron-driven Claude tasks for GitHub ops.

ai-workflows

Reusable GitHub Copilot agentic workflows. Repo-level agent recipes.

raycast-smart-issue

Raycast extension. AI-drafted GitHub issues via a local MLX model.

nix-ai

The Nix package layer that ships every AI CLI to your shell.
See AI development pipeline for how these fit into the issue → PR → ship flow.