Intent Layer

Drops AGENTS.md files throughout your codebase so Claude can navigate it like a senior engineer would. Detects what's already there, measures token counts per directory, then either scaffolds the initial structure or helps maintain existing nodes. The workflow is gated: it shows you complexity metrics before touching anything, asks whether you want CLAUDE.md or AGENTS.md at root (never both), and only creates child nodes when directories exceed 20k tokens or have clear responsibility boundaries. Includes scripts for detection and analysis, plus templates and a capture protocol for interviewing domain experts. Honest take: this is infrastructure work that pays off once your repo gets complex enough that context window management matters more than just dumping everything in.

skills.sh installs: 638
GitHub stars: 115
last updated: Aug 21, 2026

View on GitHub

Install

$ npx -y skills add crafter-station/skills --skill intent-layer --agent claude-code

Installs into .claude/skills of the current project.

Files

Intent Layer

Hierarchical AGENTS.md infrastructure so agents navigate codebases like senior engineers.

Core Principle

Only ONE root context file. CLAUDE.md and AGENTS.md should NOT coexist at project root. Child AGENTS.md in subdirectories are encouraged for complex subsystems.

Workflow

1. Detect state
   scripts/detect_state.sh /path/to/project
   → Returns: none | partial | complete

2. Route
   none/partial → Initial setup (steps 3-5)
   complete     → Maintenance (step 6)

3. Measure [gate - show table first]
   scripts/analyze_structure.sh /path/to/project
   scripts/estimate_tokens.sh /path/to/each/source/dir

4. Decide
   No root file  → Ask: CLAUDE.md or AGENTS.md?
   Has root file → Add Intent Layer section + child nodes if needed

5. Execute
   Use references/templates.md for structure
   Use references/node-examples.md for real-world patterns
   Validate: one root, READ-FIRST directive, <4k tokens per node

6. Maintenance mode (when state=complete)
   Ask user:
   a) Audit nodes     → Use references/capture-protocol.md for SME questions
   b) Find candidates → Re-measure tokens, suggest new nodes
   c) Both