agent-init-deep

Init Deep — Progressive Disclosure CLAUDE.md

Set up or migrate to a progressive disclosure CLAUDE.md structure using docs/agents/ for topic-specific guidance.

Context Model

Static (root CLAUDE.md)      — loaded every conversation, minimal, high-value
Semi-dynamic (docs/agents/)  — linked from root, loaded when the link is followed
Fully dynamic (skills)       — triggered by metadata match, loaded only when invoked

Root CLAUDE.md must be 40-50 lines maximum. Everything else belongs in docs/agents/ or skills.

Target Structure

CLAUDE.md                        # Root: identity, tech stack, key rules, workflow, links
docs/agents/
├── tooling.md                   # e.g. package manager, linting, formatting, hooks
├── commands.md                  # e.g. script execution, build filters, passing args
├── guardrails.md                # e.g. data isolation, secrets, library docs
├── definition-of-done.md       # e.g. coverage, lint, type-check, format requirements
└── [topic].md                   # Additional topic-specific files as needed

Workflow

Step 1: Detect State

Check for:
- CLAUDE.md exists?
- docs/agents/ exists?
- How many lines is CLAUDE.md?

If no CLAUDE.md → Greenfield path If CLAUDE.md exists → Migration path

Step 2a: Greenfield Path

Ask the user:

1. Project name and one-line description
2. Tech stack (frontend, backend, database, etc.)
3. Package manager (pnpm, npm, yarn, bun)
4. Key guardrails (multi-tenancy, secrets, etc.)
5. Definition of done (test coverage, lint, types, format)

Then generate:

Root CLAUDE.md with:

• Project identity (1-2 lines)
• Tech stack list
• 3-5 key rules (only things the agent consistently gets wrong)
• 4-stage workflow: Plan → Execute → Validate → Commit
• Links to docs/agents/ files with routing signals

docs/agents/ files based on answers. Common files include:

• tooling.md — package manager rules, linting config, hooks
• commands.md — how to run scripts, filter by package, pass args
• guardrails.md — data isolation, secrets, library docs
• definition-of-done.md — specific thresholds and commands

Adjust filenames and topics to match the project's actual needs.

Step 2b: Migration Path

1. Read existing CLAUDE.md

2. Classify each section:

   • Root-worthy: identity, tech stack, key rules (3-5 max), workflow
   • docs/agents/: detailed tooling, commands, guardrails, definition of done
   • Skill-worthy: complex workflows, procedures, domain expertise

3. Present proposed split to user:

   ROOT CLAUDE.md:
   - Project identity
   - Tech stack
   - Key rules: [list]
   - Workflow (Plan/Execute/Validate/Commit)
   - Links to docs/agents/
   
   docs/agents/tooling.md:
   - [extracted sections]
   
   docs/agents/commands.md:
   - [extracted sections]
   
   ... etc
   

4. Ask user to confirm or adjust

5. Create docs/agents/ files

6. Rewrite root CLAUDE.md

Step 3: Post-Setup

After creating the structure:

1. List all created files
2. Show the root CLAUDE.md
3. Suggest additional improvements:
   • "Consider adding CLAUDE.md files in app subdirectories for app-specific rules"
   • "Run /agent-add-rule to add new rules to the right location"
   • "Run /skills to see available skills"

Root CLAUDE.md Template

# Project Context

[One-line project description]

## Tech Stack

- [list technologies]

## Key Rules

- [3-5 rules the agent consistently gets wrong without being told]

## Workflow

Every task follows four stages. Identify which stage you're in and follow its rules.

Plan → Execute → Validate → Commit
↑                               |
└── fix ────────────────────────┘

1. **Plan** — Understand the task, research code, design approach. Be concise; list unresolved questions.
2. **Execute** — Implement changes AND write tests together. No implementation is complete without tests.
3. **Validate** — ALL checks must pass with zero errors before moving on:
   - [list validation commands]
     If ANY check fails → return to Execute, fix, re-validate. Pre-existing errors are NOT exempt.
4. **Commit** — Only after Validate passes completely. Never commit with failing checks.

## Detailed Guidance

When working on tasks involving these topics, read the linked doc:

- Topic (`docs/agents/file.md`) — brief routing signal describing when to read this
- Run `/skills` to see available patterns and workflows

Classification Heuristic

When deciding what stays in root vs moves to docs/agents/:

Criteria	Root	docs/agents/
Agent gets wrong without it?	YES	Sometimes
Applies to every task?	YES	NO
Under 2 lines?	YES	Any length
Detailed reference?	NO	YES
Procedural/workflow?	only the 4-stage loop	YES (multi-step)

Principles

• Minimal root: Every line in root costs tokens on every conversation. Include only rules the agent consistently gets wrong without being told.
• Routing signals: Each link description must tell Claude exactly when to read it. Be specific: "pnpm conventions, ESLint config" not just "tooling".
• One level deep: All docs link from root. No cross-references between docs/agents/ files.
• docs/agents/ not docs/: The agents/ subdirectory separates agent instructions from human documentation.

Examples

Positive Trigger

User: "Refactor our bloated CLAUDE.md into progressive disclosure with docs/agents."

Expected behavior: Use agent-init-deep guidance, follow its workflow, and return actionable output.

Non-Trigger

User: "Review this React component for unnecessary re-renders."

Expected behavior: Do not prioritize agent-init-deep; choose a more relevant skill or proceed without it.

Troubleshooting

Skill Does Not Trigger

• Error: The skill is not selected when expected.
• Cause: Request wording does not clearly match the description trigger conditions.
• Solution: Rephrase with explicit domain/task keywords from the description and retry.

Guidance Conflicts With Another Skill

• Error: Instructions from multiple skills conflict in one task.
• Cause: Overlapping scope across loaded skills.
• Solution: State which skill is authoritative for the current step and apply that workflow first.

Output Is Too Generic

• Error: Result lacks concrete, actionable detail.
• Cause: Task input omitted context, constraints, or target format.
• Solution: Add specific constraints (environment, scope, format, success criteria) and rerun.