Meta

Observations and design work toward a knowledge base for design history.

Navigation: directory index, tags index.

Goal

Build a knowledge base that applies deploy-time learning, constraining, and the generator/verifier pattern to managing design notes, decisions, and architecture.

Constraint: Claude Code as runtime

The knowledge base runs on Claude Code — using skills, hooks, and CLAUDE.md as the execution substrate. The implementation is markdown files, ripgrep queries, shell scripts, and skill definitions.

Approach

arscontexta is our first large experiment. These observations evaluate what works and inform what comes next:

• What to keep — machinery that earns its complexity (e.g., /connect for finding relationships)
• What to simplify — overhead that doesn't pay for itself (e.g., queue management, pipeline chaining)
• What to build — automated quality checks as they become justified by real failures, not taxonomy

The verifiability gradient applies to the knowledge base itself: 1. Start soft — LLM writes and connects notes (stochastic) 2. Add filters — automated checks reject bad samples (deterministic code where possible, LLM rubrics where needed) 3. Constrain search — recurring queries become indexes, tags, structured rg patterns 4. Constrain the filters — LLM rubrics that prove reliable get replaced with deterministic checks

Status

We're early. Record what you notice. Separate observations from prescriptions — note that something adds complexity before concluding what should replace it.