# Part 5 — Encoding Domain Knowledge: Asset Templates

Templates for getting your organization's knowledge *into* the agents, from the
QRefAI *AI Coding Field Guide*, Part 5. Copy into your platform/repo and adapt.

```
.
├── AGENTS.md                              # 1. Canonical AGENTS.md (blank template, fully annotated)
│                                          #    8 community-converged sections, version-pin + show-don't-tell + lean guidance
├── trio/                                  # 2. Three-vendor instruction trio (the thin-adapter pattern)
│   ├── AGENTS.md                          #    canonical SOURCE (filled worked example)
│   ├── CLAUDE.md                          #    thin adapter: @imports AGENTS.md + Claude-only overrides
│   └── copilot-instructions.md            #    thin adapter: dense first-4K restatement, then defers
├── skills/                                # 3. Starter skill library (SKILL.md stubs, description-as-trigger)
│   ├── code-review/                       #    cross-cutting
│   ├── secure-coding-checklist/
│   ├── db-migration/
│   ├── pii-handling/
│   ├── regulatory-logging/
│   ├── incident-runbook/
│   ├── payments/                          #    domain stub — fill in for your org
│   ├── kyc/                               #    domain stub
│   └── clinical-data/                     #    domain stub
└── mcp/                                   # 4. AST/graph code-intelligence MCP selection
    ├── ast-graph-mcp-selection.csv        #    comparison matrix (forkable)
    └── SELECTION-NOTES.md                 #    how to read the numbers + decision shortcuts
```

## The three ideas this page operationalizes

1. **One canonical file, thin adapters.** `AGENTS.md` is the open-standard source of
   truth; `CLAUDE.md` and `copilot-instructions.md` are thin wrappers over it. The
   `trio/` folder is a worked example you can read top-to-bottom to see exactly what
   stays canonical and what becomes a per-vendor override. Maintain content once.

2. **Lean instructions + a thick skill library.** The canonical `AGENTS.md` is
   annotated with the two rules that decide whether it helps or hurts: keep it lean
   (ETH Zurich, arXiv:2602.11988 — bloat lowers success, raises cost) and show real
   3–10 line code snippets instead of prose (lifts code-reuse ~20%). Everything
   situational moves into the `skills/` library, each stub written description-first
   so the agent loads it only when it applies.

3. **Stop grepping; query a graph.** The `mcp/` sheet compares the AST/graph
   code-intelligence servers that replace grep/glob/read exploration with single
   structured calls. Read `SELECTION-NOTES.md` before quoting any benchmark — only
   CodeGraph's numbers come from a published, dated benchmark; the rest are
   author-reported.

## Notes

- **Two `AGENTS.md` files, on purpose.** The top-level one is a *blank annotated
  template* to fill in; `trio/AGENTS.md` is a *filled worked example* showing the
  adapter relationship. Don't ship both to the same repo — pick the blank, fill it.
- **Domain skills (payments/kyc/clinical-data) are stubs**, deliberately. They carry
  the right structure and trigger pattern but must be completed with your real,
  regulation-specific rules; the placeholders say so inline.
- **Skill bodies stay lean**; each notes where bulky tables/scripts go in
  `references/` and `scripts/` (tier-3, 0 tokens until opened) — consistent with the
  progressive-disclosure model from Part 1 and the skill template in Part 2.

_Snapshot of the mid-2026 consensus. Re-verify vendor specifics — and especially the
MCP benchmark numbers — before relying on them._