[catcam]

  We ran into this exact problem and ended up formalizing what we were doing into a small convention called HADS. Four
  block types in plain Markdown — [SPEC] for authoritative facts, [NOTE] for context, [BUG] for known failures, [?] for
  unverified — plus an AI manifest at the top that tells the model what to read and skip. No tooling, just annotation.

  In practice it cut per-query token load ~70% on our longer docs. Small models (7B) handle it well because the tags
  remove the structural reasoning problem entirely.

  Spec and examples: https://github.com/catcam/hads