Two Tier Docs Pattern

Two-Tier Docs Pattern

Keep root context files short; put deep references in a separate directory. The two tiers serve different purposes and have different cost models. Conflating them produces docs that fail at both jobs.

For the canonical version, see agentic-docs/two-tier-docs-pattern.md.

The pattern

Tier 1: Root context files                Tier 2: Deep reference library
(loaded into every conversation)          (read on demand by humans and agents)

AGENTS.md         (~100 lines)            agentic-docs/
CLAUDE.md         (~50 lines, shim)       ├── philosophy.md
GEMINI.md         (~30 lines, shim)       ├── spec-driven-development.md
                                          ├── agentic-coding-model.md
                                          ├── automation-decision-framework.md
                                          ├── ...

Tier 1 is short, scannable, and load-bearing for every interaction. Tier 2 is detailed, comprehensive, and read selectively on demand.

AGENTS.md standard vs SpecRoute convention. The Agentic AI Foundation standard is plain Markdown with no required fields, frontmatter, schema, or official length cap; nested files use nearest-file precedence. The ~100-line target below is SpecRoute's context-budget convention, not an external requirement.

Why two tiers

The cost of long root context

Every file the agent CLI loads at session start consumes context budget. A 2,000-line root context file is a slow leak from the agent's working memory. It also makes orientation slower — the user wanting "what should I do for a security incident?" doesn't need the spec-driven primer first.

The cost of fragmented short docs

If everything is split into 50-line files with cross-references, navigation becomes a chore. The reader follows three links to assemble a complete picture.

The compromise

• Root files carry the project's identity, conventions, and pointers. Optimized for "agent loads this at every session start."
• Deep references carry the substance. Optimized for "reader follows a link when they need this."

Same pattern operating systems use (kernel API surface vs. man pages), and the open-source world has converged on (README.md for orientation, agentic-docs/ for substance).

Tier 1: what belongs

In AGENTS.md (canonical) and per-vendor delegation shims:

• What this repository is — one paragraph.
• Core artifact taxonomy — a table.
• Hard constraints — the rules that aren't obvious from the code.
• Pointers to deep references — "for X, see agentic-docs/<topic>.md".
• Vendor-specific overrides (in the per-vendor shims only).

Tier 2: what belongs

In agentic-docs/:

• Philosophy — why we do things the way we do.
• Decision frameworks — when to choose X vs Y.
• Patterns — recurring shapes.
• Integration guides — concrete wiring per vendor.
• Conceptual deep-dives — anything substantive that doesn't fit in 30 lines.

How the tiers reference each other

Tier 1 → Tier 2: pointers only.

# AGENTS.md
When to choose a skill vs an agent vs a command vs a hook: see
[`agentic-docs/automation-decision-framework.md`](agentic-docs/automation-decision-framework.md).

Tier 2 → Tier 1: assume the reader has read it. No need to re-explain the artifact taxonomy.

Sizing rules of thumb

Tier	Target	Hard cap
AGENTS.md	~100 lines	~150 lines
CLAUDE.md, GEMINI.md (delegation shims)	30–50 lines	75 lines
agentic-docs/<topic>.md	100–300 lines	flexible

If a Tier 1 file passes the hard cap, ask which sections can move to Tier 2.

What goes wrong when you collapse tiers

• Single huge root file → slow context loading, hard to scan, agent response quality degrades.
• Tier 2 only (no root file) → agent CLIs have no anchor; conventions get re-explained in every conversation.
• Tier 1 with substantive content → root file balloons; concepts duplicated between tiers; updates require multi-file edits.

Vendor-specific notes

• Claude Code loads CLAUDE.md automatically. Keep it short; deep references in agentic-docs/ are loaded on demand.
• Codex loads AGENTS.md automatically.
• Gemini CLI loads GEMINI.md automatically.
• Cursor / Devin Desktop read AGENTS.md natively and also load rule files (.cursor/rules/*.mdc, .devin/rules/*.md; Cascade additionally accepts .windsurf/rules/). The two-tier pattern is enforced through always-loaded versus scoped rules.
• Kiro loads .kiro/steering/*.md per inclusion frontmatter.

See Multi-Vendor Context Files for the canonical-shim-pattern in full.

See also

• Multi-Vendor Context Files — how the canonical AGENTS.md and per-vendor shims work
• Documentation Structure — where new docs go
• Rules — content standards including documentation rules