https://github.com/catcam/hads
AI models increasingly read documentation before humans do. But docs are written for humans — verbose, contextual, narrative. This creates token waste and increases hallucination risk, especially on smaller/local models.
HADS is not a new format. It's a tagging convention on top of standard Markdown:
[SPEC] — authoritative facts, terse, bullet/table/code
[NOTE] — human context, history, examples
[BUG] — verified failure + fix (symptom, cause, fix)
[?] — unverified/inferred, lower confidence
A 7B local model with limited context window can read a HADS document and extract facts correctly because it doesn't have to reason about structure — the document tells it how.
The repo includes: - Full specification (SPEC.md) - Three example documents (REST API, binary file format, config system) - Python validator (exit codes for CI/CD) - Claude skill (SKILL.md) for AI-assisted doc generation
All MIT. Feedback welcome — especially from people running local models.
Rough Fermi estimate of global impact if this became standard practice:
Assumptions:
- ~1B AI queries/day touch technical docs (Copilot, ChatGPT, enterprise agents, CI/CD pipelines)
- HADS reduces tokens read per query from ~5k to ~1.5k via manifest + targeted block reading (70% reduction)
- ~0.003 Wh per 1k tokens for GPT-4 class inference (per published ML efficiency research)
- ~2 min saved per query from more precise first answers
Results:
- ~3.8 TWh electricity saved/year (~350k US homes, or ~20% of Google's annual data center consumption)
- ~12B developer hours recovered/year
- ~$60B/year in productivity at 10% adoption
These are intentionally rough — could be 10x off in either direction. The point is that documentation structure is
economically material at scale, not just an ergonomics concern.
Full writeup with methodology:
https://medium.com/@catcam_46604/ai-is-now-the-primary-reader-of-your-docs-nobody-told-your-docs-5f7103ea3281