ai-first-documentation

name

ai-first-documentation

description

Research-backed standards for documentation in AI-first codebases: context-file architecture (CLAUDE.md, skills, subagents, rules), multi-agent interop, and project-doc organization, with a tiered audit rubric. Use when authoring or auditing CLAUDE.md/AGENTS.md, deciding where guidance belongs, or running a documentation gap analysis. DO NOT USE FOR: post-implementation doc updates (use documentation-finalization) or reference sidecar setup (use project-references).

AI-First Documentation Standards

Standards for structuring a repository's documentation when coding agents are primary readers and operators. Every practice here is backed by two adversarially-verified research passes (June 2026) across Anthropic official guidance (Tier 1), major vendor guidance (Tier 2), and evidenced practitioner sources (Tier 3). The testable rubric lives in rubric.md; citations, verification dates, refuted claims, and open gaps live in sources.md.

• When to Use
• Core Principle
• The Placement Model
• Audit Workflow
• Consumer-Mode Audits
• Recording Documentation Decisions
• Multi-Agent Repositories
• Known Industry Open Gaps
• Quick Reference — Verified Numbers
• See Also
• Gotchas

When to Use

• Authoring or auditing a CLAUDE.md, AGENTS.md, or other always-loaded context file
• Deciding where a piece of guidance belongs (always-loaded vs. path-scoped vs. skill vs. plain doc)
• Designing or reviewing skills, subagent definitions, or rules files
• Running a documentation gap analysis of a repository against industry best practice
• Configuring a repo that must serve multiple agent toolchains (Claude Code, Copilot, Cursor, Codex)

Core Principle

The context window is the scarce resource. Performance degrades as it fills, so always-loaded files must be ruthlessly minimal — every line survives the deletion test ("Would removing this cause the agent to make mistakes?") — while everything else moves to on-demand mechanisms: skills with progressive disclosure, path-scoped rules, child-directory context files, and just-in-time retrieval via glob/grep rather than pre-built indexes. (Tier 1; corroborated by all surveyed vendors and an academic taxonomy of 13 open-source agents.)

Minimal does not mean short for its own sake: the criterion is signal density plus behavioral sufficiency, not raw line count.

For the model-generation dimension of this principle — separating load-bearing contracts from Claude-4-era behavioral coaching when auditing imperative-guardrail density — see Documents/Design/context-engineering-claude-5.md.

The Placement Model

Route each piece of guidance to its cheapest sufficient home:

Content kind	Home	Loading cost
Broadly-applicable facts: non-guessable commands, non-default conventions, testing instructions, repo etiquette, gotchas, "always do X" rules	Root CLAUDE.md / AGENTS.md	Every session
Area-specific instructions tied to file paths	Path-scoped rules (paths:/glob frontmatter) or child-directory CLAUDE.md	Only when matching files are touched
Multi-step procedures, domain workflows, methodology	Skill (SKILL.md)	Metadata always (~100 tokens); body on demand
Deep reference material, schemas, long examples	Skill reference files or plain docs, pointed to one level deep	Zero until accessed
Side-effectful workflows that must not auto-trigger	Skill with disable-model-invocation: true	Manual invocation only
High-volume exploration or review output	Subagent with isolated context	Never enters main context
Anything derivable by reading the code, standard language conventions, frequently-changing facts, file-by-file codebase tours	Nowhere — delete it	—

Two traps in this model: @imports in CLAUDE.md load at launch and save no context (only child-directory files and skills defer cost), and rules files without a path scope load unconditionally every session.

Audit Workflow

To run a gap analysis of a repository:

1. Inventory the documentation surface: root and nested context files (CLAUDE.md, AGENTS.md, CLAUDE.local.md), .claude/rules/, skills/*/SKILL.md, agent definitions, and the plain-doc tree (design docs, ADRs, runbooks, plans).
2. Classify each artifact against the Placement Model: is anything always-loaded that could be on-demand? Is anything on-demand that the agent needs every session?
3. Test each applicable practice in rubric.md, section by section (A: always-loaded context, B: skills, C: subagents, D: writing style, E: multi-agent interop, F: project docs).
4. Weight findings by tier and confidence from the rubric: a Tier 1 high-confidence violation is a defect; a Tier 2 conflict is a trade-off to decide consciously; a Tier 3 divergence is informational.
5. Label house conventions in open-gap areas. Where the rubric marks a topic as a confirmed industry open gap (Section F), a repo's own conventions are not violations — record them as house practice filling a documented gap, and consider them candidates for promotion if external standards later emerge.

The mechanical checks (A2, B2, B3, B5, A9) are deterministic and can be run headlessly via scripts/audit-docs-mechanical.ps1 with an explicit -Root parameter. For the interactive audit experience (applying judgment, browsing findings, running init), use /audit-docs — it runs the script and applies rubric judgment passes on top.

Consumer-Mode Audits

The mechanical-check script (scripts/audit-docs-mechanical.ps1) is designed to run against any consumer repository root, not just the hub. When run from a consumer's CI or shell:

• Own-surface scoping rule (D3): the script inventories only files in the consumer tree — root CLAUDE.md, AGENTS.md, CLAUDE.local.md, .claude/rules/, local skills/, .claude/agents/, and the project doc tree. Plugin-cache paths (.claude/plugins/) are never inventoried.
• Entry-point precedence (D7): for interactive audits, /audit-docs is the deterministic entry point — it runs the script and applies rubric judgment. For headless/CI use, invoke the script directly:

  pwsh skills/ai-first-documentation/scripts/audit-docs-mechanical.ps1 -Root . -FailOn fail
  

  Acquire the script via vendor-copy or pinned raw fetch; do not depend on a git submodule or npm package.
• CI snippet (acquire + run):

  # After vendoring the script to your repo (e.g., scripts/audit-docs-mechanical.ps1):
  pwsh scripts/audit-docs-mechanical.ps1 -Root $env:GITHUB_WORKSPACE -FailOn fail
  

  Always pass a non-empty -Root — an empty string produces a PowerShell parameter-binding error rather than JSON output, since the binding check runs before the script body. The command layer's root-resolution step (CWD fallback) prevents this in interactive use.

Recording Documentation Decisions

When a check fires and the deviation is intentional (e.g., CLAUDE.md is temporarily over 200 lines during a migration), record the decision in .claude/documentation-decisions.md at the consumer's root. The script reads this file and emits status: waived instead of fail for matched entries, so the decision is not re-litigated on re-runs or in CI.

Format (named decision d-decision-record-format) — H3-per-record markdown:

### {check-id}: {relative-path}
- rationale: [CUSTOMIZE: why this deviation is intentional and temporary]
- date: YYYY-MM-DD

Example:

### A2: CLAUDE.md
- rationale: CLAUDE.md is intentionally long during the Q3 migration; pruning tracked in issue #42.
- date: 2026-06-12

The match key is (check_id, normalized relative path from root). The waiver_ref in the JSON output is the raw heading text (A2: CLAUDE.md). Keep the H3 heading to the two-token {check-id}: {relative-path} form — any trailing annotation (e.g., ### A2: CLAUDE.md — migration note) will appear verbatim in waiver_ref. Matching is case-insensitive: ### a2: claude.md matches the same entry as ### A2: CLAUDE.md.

Relocation: the default location is .claude/documentation-decisions.md. Override with -DecisionRecordPath <absolute-or-repo-relative-path> when your repo layout differs.

Multi-Agent Repositories

When one repo serves several agent toolchains, consult rubric Section E before restructuring. Key verified facts: AGENTS.md is the cross-vendor de facto standard, but Claude Code does not read it natively (use a CLAUDE.md containing @AGENTS.md); Copilot's cloud agent reads CLAUDE.md natively but only as a single flat root file; Copilot code review reads only the first 4,000 characters of any instruction file; Cursor silently ignores plain .md files in .cursor/rules (frontmatter .mdc or AGENTS.md only); Codex concatenates AGENTS.md files root-down with the closest file winning positionally.

One genuine unresolved conflict exists: GitHub recommends embedding repo structure and dependency versions and warns that pointer references to other docs may be ineffective in large repos — the opposite of Anthropic's deletion-test minimalism and progressive disclosure. Neither side has measured evidence. Default to the Anthropic model for Claude-primary repos; consciously revisit if Copilot adherence problems appear.

Known Industry Open Gaps

No credible verified guidance exists (as of June 2026) for: authoring ADRs/design docs differently for AI readers; whether llms.txt-style index files help or hurt vs. pure just-in-time search; wrapping general project docs as skills; doc-to-code linking that breaks loudly on drift; or doc-tree naming conventions optimized for agent navigation. The only verified adjacent practices are staleness automation (review-time flagging of doc drift — "a stale context file can be worse than no file at all") and the split-and-reference pattern (concise root file pointing to task-specific docs). Treat conventions in these areas as labeled house judgment, not settled standards. Details in sources.md.

Quick Reference — Verified Numbers

Number	What it bounds	Tier
< 200 lines	CLAUDE.md adherence target (not a truncation limit)	1
< 500 lines	SKILL.md body; also Cursor's per-rule-file ceiling	1 / 2
64 / 1024 chars	Skill frontmatter name / description hard limits	1
1 level	Skill reference-file depth from SKILL.md (never nested chains)	1
> 100 lines	Reference file size that requires a leading table of contents	1
Depth 4	Maximum recursive @import depth in CLAUDE.md	1
4,000 chars	Copilot code review reads only this much of any instruction file	2
32 KiB	Codex default combined instruction-file budget (project_doc_max_bytes)	2

All numbers live-verified 2026-06-10/11; vendor docs churn, so re-verify before hard enforcement (see sources.md).

See Also

• rubric.md — the practice-by-practice testable audit rubric
• sources.md — citations, verification dates, refuted claims, open questions
• scripts/audit-docs-mechanical.ps1 — mechanical-check script for A2, B2, B3, B5, A9; headless/CI mode
• templates/CLAUDE.md-starter.md — minimal CLAUDE.md seed template
• documentation-finalization — updating implementation-facing docs after code changes
• project-references — Agent Orchestra reference sidecars and index conventions (a house convention in open-gap territory)
• skill-creator — repo-specific skill frontmatter and structure conventions

Gotchas

Trigger	Gotcha	Fix
Moving CLAUDE.md overflow into @import files "to save context"	Imports load in full at launch — zero context savings, same adherence cost	Move content to a skill or child-directory CLAUDE.md, which load on demand
Adding a rules file without a path scope	It loads every session, silently becoming always-loaded context	Add a paths:/glob scope, or move the content to a skill
Writing a skill description in first person or omitting "when to use"	Description is the sole selection signal injected into the system prompt; the skill is never picked	Third person, state what it does AND when to use it, within 1024 chars
Chaining reference files (SKILL.md → ref-a.md → ref-b.md)	Agents may partially read nested references (e.g., head-100 previews) and miss content	Link every reference file directly from SKILL.md, one level deep
Assuming Copilot honors nested CLAUDE.md or @-imports	Copilot's cloud agent reads only a single flat root CLAUDE.md	Keep the root file self-sufficient for Copilot, or maintain AGENTS.md
Padding CLAUDE.md with folder structure and dependency versions	JIT-discoverable, staleness-prone content that fails the deletion test (GitHub recommends it; Anthropic canon excludes it — known conflict)	Default to omitting; the agent can glob/grep it fresh