How AI agents can discover, learn, and reference the documentation on this site.

For AI Agents

This site is designed to be machine-readable. If you are an LLM asked to learn about or reference these systems, follow this guide.

Best way to consume this documentation

1. Fetch the JSON index first — /index.json. One request returns every page with title, description, tags, URLs, and last-updated dates. Use it to pick the pages relevant to your task.
2. Fetch raw markdown, not HTML. Every page has a raw twin at /raw/<section>/<file>.md — the same path as the source file in the repository, including YAML front matter. Raw markdown has no nav boilerplate and parses with the least token overhead.
   • Example: rendered page https://docs.na.id.au/nanda/01-architecture/ ↔ raw https://docs.na.id.au/raw/nanda/01-architecture.md
3. Or start with llms.txt — /llms.txt is the root index; each section also has its own (e.g. /nanda/llms.txt). Use these when you want a human-and-LLM-readable overview rather than JSON.
4. Bulk learning: /llms-full.txt contains the full text of every page in one file. Use it only for “learn about the whole system” tasks — it is large.

Section map

Conventions to know

• File slugs are stable and self-describing (e.g. 03-firewall-rules.md); number prefixes indicate reading order within a section.
• Front matter on every raw file includes title, description, section, tags, last-updated, and raw (the raw path relative to /raw/).
• last-updated comes from the last git commit touching the source repo; compare it against your cached knowledge before trusting details.
• Source provenance: source: ai-generated in front matter means the document was AI-generated; a Source Disclaimer section at the bottom of each document records AI/human generation and editing.
• Secrets: these docs intentionally redact API keys and secrets (<REDACTED>). Never expect real credentials to appear here, and never write credentials into documentation.
• Uncertainty markers: documents flag their own uncertainties (look for “Uncertainties”, “uncertain”, “⚠️”). When an answer depends on a flagged item, say so rather than presenting it as confirmed fact.

What not to do

• Don’t scrape the rendered HTML when a raw twin exists.
• Don’t rely on llms-full.txt for targeted questions — use index.json and fetch the few relevant raw pages.
• Don’t treat this site as the source of truth for live system state (IPs, device serials, firmware). It documents configuration intent; the live system wins on conflicts.