Principles

Apply these throughout the pipeline below; they decide what to read, what to cache, and what to skip.

Optimize for token efficiency

• Process documents (code, configuration, data, prose) with sub-agents, not in one long context.
• Prefer programmatic extraction and classification over reading full document contents.
• Store SHA256 hashes to detect changes programmatically, enabling caching of prior results.
• Generate and reuse utility scripts to avoid duplicate AI inference.
• Split work into sub-tasks that fit comfortably under the model's context limit (≈200k tokens). Exceptions allowed.

Don't reinvent the wheel

The domain may already have established tools to map system state, index code, or generate a documentation tree. Evaluate whether they fit before rolling your own — building on an existing, well-tested solution is usually better unless it needs significant changes or has unacceptable limitations.

Incremental by default