Deep Dive Analysis Skill

Overview

This skill combines mechanical structure extraction with Claude's semantic understanding to produce comprehensive codebase documentation. Unlike simple AST parsing, this skill captures:

WHAT the code does (structure, functions, classes)
WHY it exists (business purpose, design decisions)
HOW it integrates (dependencies, contracts, flows)
CONSEQUENCES of changes (side effects, failure modes)

Capabilities

Mechanical Analysis (Scripts):

Extract code structure (classes, functions, imports)
Map dependencies (internal/external)
Find symbol usages across the codebase
Track analysis progress
Classify files by criticality

Semantic Analysis (Claude AI):

Recognize architectural and design patterns
Identify red flags and anti-patterns
Trace data and control flows
Document contracts and invariants
Assess quality and maintainability

Documentation Maintenance:

Review and maintain documentation (Phase 8)
Fix broken links and update navigation indexes
Analyze and rewrite code comments (antirez standards)

Use this skill when:

Analyzing a codebase you're unfamiliar with
Generating documentation that explains WHY, not just WHAT
Identifying architectural patterns and anti-patterns
Performing code review with semantic understanding
Onboarding to a new project
Creating documentation for new contributors

Prerequisites

analysis_progress.json must exist in project root (created by DEEP_DIVE_PLAN setup)
DEEP_DIVE_PLAN.md should be reviewed to understand phase structure

CRITICAL PRINCIPLE: ABSOLUTE SOURCE OF TRUTH

THE DOCUMENTATION GENERATED BY THIS SKILL IS THE ABSOLUTE AND UNQUESTIONABLE SOURCE OF TRUTH FOR YOUR PROJECT.

ANY INFORMATION NOT VERIFIED WITH IRREFUTABLE EVIDENCE FROM SOURCE CODE IS FALSE, UNRELIABLE, AND UNACCEPTABLE.

IMPORTANT LIMITATION: Verification is Multi-Layer

╔══════════════════════════════════════════════════════════════════════════════╗
║                     VERIFICATION TRUST MODEL                                  ║
╠══════════════════════════════════════════════════════════════════════════════╣
║  Layer 1: TOOL-VALIDATED                                                     ║
║    └── Automated checks: file exists, line in range, AST symbol match        ║
║    └── Marker: [VALIDATED: file.py:123 @ 2025-12-20]                         ║
║                                                                              ║
║  Layer 2: HUMAN-VERIFIED                                                     ║
║    └── Manual review: semantic correctness, behavior match                   ║
║    └── Marker: [VERIFIED: file.py:123 by @reviewer @ 2025-12-20]             ║
║                                                                              ║
║  Layer 3: RUNTIME-CONFIRMED                                                  ║
║    └── Log/trace evidence of actual behavior                                 ║
║    └── Marker: [CONFIRMED: trace_id=abc123 @ 2025-12-20]                     ║
╚══════════════════════════════════════════════════════════════════════════════╝

Tool validation catches STRUCTURAL issues (file moved, line shifted, symbol renamed).
Human verification ensures SEMANTIC correctness (code does what doc says).
Runtime confirmation proves BEHAVIORAL truth (system actually works this way).

ALL THREE LAYERS are required for critical documentation.

The Iron Law of Documentation

╔══════════════════════════════════════════════════════════════════════════════╗
║  DOCUMENTATION = f(SOURCE_CODE) + VERIFICATION                               ║
║                                                                              ║
║  If NOT verified_against_code(statement) → statement is FALSE                ║
║  If NOT exists_in_codebase(reference)    → reference is FABRICATED           ║
║  If NOT traceable_to_source(claim)       → claim is SPECULATION              ║
╚══════════════════════════════════════════════════════════════════════════════╝

Mandatory Rules (VIOLATION = FAILURE)

NEVER document anything without reading the actual source code first
NEVER assume any existing documentation, comment, or docstring is accurate
NEVER write documentation based on memory, inference, or "what should be"
ALWAYS derive truth EXCLUSIVELY from reading and tracing actual code
ALWAYS provide source file + line number for every technical claim
ALWAYS verify state machines, enums, constants against actual definitions
TREAT all pre-existing docs as unverified claims requiring validation
MARK any unverifiable statement as [UNVERIFIED - REQUIRES CODE CHECK]

Verification Requirements

Documentation Type	Required Evidence
Enum/State values	Exact match with source code enum definition
Function behavior	Code path tracing, actual implementation reading
Constants/Timeouts	Variable definition in source with file:line
Message formats	Message class definition, field validation
Architecture claims	Import graph analysis, actual class relationships
Flow diagrams	Verified against runtime logs OR code path analysis

Documentation Verification Status

Every section of documentation MUST have one of these status markers:

[VERIFIED: source_file.py:123] - Confirmed against source code
[VERIFIED: trace_id=xyz] - Confirmed against runtime logs
[UNVERIFIED] - Requires verification before trusting
[DEPRECATED] - Code has changed, documentation outdated

UNVERIFIED documentation is UNTRUSTED documentation.

CRITICAL PRINCIPLE: NO HISTORICAL DEPTH

DOCUMENTATION DESCRIBES ONLY THE CURRENT STATE OF THE ART.

NO HISTORY. NO ARCHAEOLOGY. NO "WAS". ONLY "IS".

╔══════════════════════════════════════════════════════════════════════════════╗
║                     THE TEMPORAL PURITY PRINCIPLE                            ║
╠══════════════════════════════════════════════════════════════════════════════╣
║  Documentation = PRESENT_TENSE(current_implementation)                       ║
║                                                                              ║
║  FORBIDDEN:                                                                  ║
║  ✗ "was/were/previously/formerly/used to"                                    ║
║  ✗ "deprecated since version X" → just REMOVE it                             ║
║  ✗ "changed from X to Y" → only describe Y                                   ║
║  ✗ "in the old system..." → irrelevant, delete                               ║
║  ✗ inline changelogs → use CHANGELOG.md or git                               ║
║                                                                              ║
║  REQUIRED:                                                                   ║
║  ✓ Present tense: "The system uses..." not "The system used..."              ║
║  ✓ Current state only: Document what IS, not what WAS                        ║
║  ✓ Git for archaeology: History lives in version control, not docs           ║
╚══════════════════════════════════════════════════════════════════════════════╝

The Rule:

When you find documentation containing historical language, DELETE IT. Git blame exists for archaeology. Documentation exists for the present.

Available Commands

1. Analyze Single File

Extract structure, dependencies, and usages for one file:

python .claude/skills/deep-dive-analysis/scripts/analyze_file.py \
  --file src/utils/circuit_breaker.py \
  --output-format markdown

Parameters:

--file / -f: Relative path to file to analyze - REQUIRED
--output-format / -o: Output format (json, markdown, summary) - default: summary
--find-usages / -u: Also find all usages of exported symbols - default: false
--update-progress / -p: Update analysis_progress.json - default: false

Output includes:

File classification (Critical/High-Complexity/Standard/Utility)
Classes with methods and attributes
Functions with signatures
Internal imports (within project)
External imports (third-party)
External calls (database, network, filesystem, messaging, ipc)
State mutations identified
Error handling patterns

2. Check Progress

View analysis progress by phase:

python .claude/skills/deep-dive-analysis/scripts/check_progress.py \
  --phase 1 \
  --status pending

Parameters:

--phase / -p: Filter by phase number (1-7)
--status / -s: Filter by status (pending, analyzing, done, blocked)
--classification / -c: Filter by classification (critical, high-complexity, standard, utility)
--verification-needed: Show only files needing runtime verification

3. Find Usages

Find all usages of a symbol across the codebase:

python .claude/skills/deep-dive-analysis/scripts/analyze_file.py \
  --symbol CircuitBreaker \
  --file src/utils/circuit_breaker.py

4. Generate Phase Report

Generate documentation for an entire phase:

python .claude/skills/deep-dive-analysis/scripts/analyze_file.py \
  --phase 1 \
  --output-format markdown \
  --output-file docs/01_domains/COMMON_LIBRARY.md

Phase 8: Documentation Review Commands

5. Scan Documentation Health

Discover all documentation files and generate health report:

python .claude/skills/deep-dive-analysis/scripts/doc_review.py scan \
  --path docs/ \
  --output doc_health_report.json

Output includes:

Total file count per directory
Files with TODO/FIXME/TBD markers
Files missing last_updated metadata
Large files (>1500 lines) candidates for splitting

6. Validate Links

Find all broken links in documentation:

python .claude/skills/deep-dive-analysis/scripts/doc_review.py validate-links \
  --path docs/ \
  --fix  # Optional: auto-remove broken links

Actions:

Extracts all relative markdown links ](../path/to/file.md)
Verifies target files exist
Reports broken links with source file and line number
With --fix: removes or updates broken references

7. Verify Against Source Code

Verify documentation accuracy against actual source code:

python .claude/skills/deep-dive-analysis/scripts/doc_review.py verify \
  --doc docs/agents/lifecycle.md \
  --source src/agents/lifecycle.py

Verification includes:

Documented states vs actual enum values
Documented methods vs actual class methods
Documented constants vs actual values
Flags discrepancies as DRIFT

8. Update Navigation Indexes

Refresh SEARCH_INDEX.md and BY_DOMAIN.md with current file counts:

python .claude/skills/deep-dive-analysis/scripts/doc_review.py update-indexes \
  --search-index docs/00_navigation/SEARCH_INDEX.md \
  --by-domain docs/00_navigation/BY_DOMAIN.md

Updates:

Total file counts
Files per directory statistics
Version and last_updated timestamps
Removes references to deleted files

9. Full Documentation Maintenance

Run complete Phase 8 workflow:

python .claude/skills/deep-dive-analysis/scripts/doc_review.py full-maintenance \
  --path docs/ \
  --auto-fix \
  --output doc_health_report.json

Executes in order:

Scan documentation health
Validate and fix broken links
Identify obsolete files (no inbound links, references deleted code)
Update navigation indexes
Generate final health report

Comment Quality Commands (Antirez Standards)

These commands analyze and rewrite code comments following the antirez commenting standards.

10. Analyze Comment Quality

Analyze comments in a single file:

python .claude/skills/deep-dive-analysis/scripts/rewrite_comments.py analyze \
  src/main.py \
  --report

Options:

--report / -r: Generate detailed markdown report
--json: Output as JSON for programmatic use
--issues-only / -i: Show only problematic comments

Output includes:

Comment classification (function, design, why, teacher, checklist, guide)
Issue detection (trivial, debt, backup comments)
Suggested rewrites for problematic comments
Statistics and ratios

11. Scan Directory for Comment Issues

Analyze all Python files in a directory:

python .claude/skills/deep-dive-analysis/scripts/rewrite_comments.py scan \
  src/ \
  --recursive \
  --issues-only

Options:

--recursive / -r: Include subdirectories
--issues-only / -i: Show only files with issues
--json: Output as JSON

12. Generate Comment Health Report

Create comprehensive markdown report for entire codebase:

python .claude/skills/deep-dive-analysis/scripts/rewrite_comments.py report \
  src/ \
  --output comment_health.md

Report includes:

Executive summary with totals
Comment quality breakdown (keep/enhance/rewrite/delete)
Comment type distribution
Files needing attention (ranked by issue count)
Sample issues with file:line references
Actionable recommendations

13. Rewrite Comments

Apply comment improvements to a file:

# Dry run (preview changes)
python .claude/skills/deep-dive-analysis/scripts/rewrite_comments.py rewrite \
  src/main.py

# Apply changes with backup
python .claude/skills/deep-dive-analysis/scripts/rewrite_comments.py rewrite \
  src/main.py \
  --apply \
  --backup

Options:

--apply / -a: Actually modify the file (default: dry run)
--backup / -b: Create .bak backup before modifying
--output / -o: Write to different file instead of in-place

Actions taken:

DELETE: Remove trivial comments and backup (commented-out code)
REWRITE: Add suggested improvements for debt comments (TODO/FIXME)

14. View Standards Reference

Display the antirez commenting standards:

python .claude/skills/deep-dive-analysis/scripts/rewrite_comments.py standards

Shows the complete taxonomy of good vs bad comments with examples.

Comment Type Classification

Type	Category	Description	Action
function	GOOD	API docs at function/class top	Keep/Enhance
design	GOOD	File-level algorithm explanations	Keep
why	GOOD	Explains reasoning behind code	Keep
teacher	GOOD	Educates about domain concepts	Keep
checklist	GOOD	Reminds of coordinated changes	Keep
guide	GOOD	Section dividers, structure	Keep sparingly
trivial	BAD	Restates what code says	Delete
debt	BAD	TODO/FIXME without plan	Rewrite/Resolve
backup	BAD	Commented-out code	Delete

Comment Quality Workflow

1. SCAN
   ├── Run: rewrite_comments.py scan <dir> --recursive
   ├── Review files with most issues
   └── Generate: rewrite_comments.py report <dir> --output report.md

2. TRIAGE
   ├── Identify high-priority files (critical modules)
   ├── Focus on DEBT comments (convert to issues or design docs)
   └── Plan bulk TRIVIAL/BACKUP deletions

3. REWRITE
   ├── Run: rewrite_comments.py rewrite <file> --apply --backup
   ├── Review changes in diff
   └── Verify no functional changes

4. VERIFY
   ├── Run tests to confirm no breakage
   ├── Re-scan to confirm improvements
   └── Update comment_health.md report

File Classification Criteria

Classification	Criteria	Verification
Critical	Handles authentication, security, encryption, sensitive data	Mandatory
High-Complexity	>300 LOC, >5 dependencies, state machines, async patterns	Mandatory
Standard	Normal business logic, data models, utilities	Recommended
Utility	Pure functions, helpers, constants	Optional

AI-Powered Semantic Analysis

This skill leverages Claude's code comprehension capabilities for deep semantic analysis beyond mechanical structure extraction.

The Semantic Analysis Mandate

╔══════════════════════════════════════════════════════════════════════════════╗
║                    STRUCTURE vs MEANING                                      ║
╠══════════════════════════════════════════════════════════════════════════════╣
║                                                                              ║
║  Scripts extract STRUCTURE:  "class Foo with method bar()"                   ║
║  Claude extracts MEANING:    "Foo implements Repository pattern for         ║
║                               caching user sessions with TTL expiration"     ║
║                                                                              ║
║  NEVER stop at structure. ALWAYS pursue understanding.                      ║
║                                                                              ║
╚══════════════════════════════════════════════════════════════════════════════╝

Five Layers of Understanding

Layer	What	Who Does It
1. WHAT	Classes, functions, imports	Scripts (AST)
2. HOW	Algorithm details, data flow	Claude's first pass
3. WHY	Business purpose, design decisions	Claude's deep analysis
4. WHEN	Triggers, lifecycle, concurrency	Claude's behavioral analysis
5. CONSEQUENCES	Side effects, failure modes	Claude's systems thinking

Semantic Analysis Questions

For every code unit, Claude must answer:

Identity:

What is this code's single responsibility?
What abstraction does it represent?
What would break if this didn't exist?

Behavior:

What are ALL inputs and outputs (including side effects)?
What state does it read? What does it mutate?
What are preconditions and postconditions?

Integration:

Who calls this? Under what circumstances?
What does this call? Why those dependencies?
What contracts does it fulfill?

Quality:

What could go wrong? How is failure handled?
Are there implicit assumptions that could break?
Are there race conditions or timing dependencies?

Pattern Recognition

Claude should actively recognize and document common patterns:

Pattern Type	Examples	Documentation Focus
Architectural	Repository, Service, CQRS, Event-Driven	Responsibilities, boundaries
Behavioral	State Machine, Strategy, Observer, Chain	Transitions, variations
Resilience	Circuit Breaker, Retry, Bulkhead, Timeout	Thresholds, fallbacks
Data	DTO, Value Object, Aggregate	Invariants, relationships
Concurrency	Producer-Consumer, Worker Pool	Thread safety, backpressure

See references/SEMANTIC_PATTERNS.md for detailed recognition guides.

Red Flags to Identify

Claude should actively flag these issues:

ARCHITECTURE:
⚠ GOD CLASS: >10 public methods or >500 LOC
⚠ CIRCULAR DEPENDENCY: A → B → C → A
⚠ LEAKY ABSTRACTION: Implementation details in interface

RELIABILITY:
⚠ SWALLOWED EXCEPTION: Empty catch blocks
⚠ MISSING TIMEOUT: Network calls without timeout
⚠ RACE CONDITION: Shared mutable state without sync

SECURITY:
⚠ HARDCODED SECRET: Passwords, API keys in code
⚠ SQL INJECTION: String concatenation in queries
⚠ MISSING VALIDATION: Unsanitized user input

Semantic Analysis Template

Use templates/semantic_analysis.md for comprehensive per-file analysis that includes:

Executive summary (purpose, responsibility, patterns)
Behavioral analysis (triggers, processing, side effects)
Dependency analysis (why each dependency exists)
Quality assessment (strengths, concerns, red flags)
Contract documentation (full interface semantics)
Flow tracing (primary and error paths)
Testing implications (what must be tested)

AI Analysis Workflow

1. SCRIPTS RUN FIRST
   ├── classifier.py → File classification
   ├── ast_parser.py → Structure extraction
   └── usage_finder.py → Cross-references

2. CLAUDE ANALYZES
   ├── Read actual source code
   ├── Apply semantic questions
   ├── Recognize patterns
   ├── Identify red flags
   └── Trace flows

3. CLAUDE DOCUMENTS
   ├── Use semantic_analysis.md template
   ├── Explain WHY, not just WHAT
   ├── Document contracts and invariants
   └── Flag concerns with severity

4. VERIFY
   ├── Check against runtime behavior
   ├── Validate with code traces
   └── Mark verification status

Reference Documents

references/AI_ANALYSIS_METHODOLOGY.md - Complete analysis methodology
references/SEMANTIC_PATTERNS.md - Pattern recognition guide
templates/semantic_analysis.md - Per-file analysis template

Analysis Loop Workflow

When analyzing a file, follow this sequence:

1. CLASSIFY
   ├── Count lines of code
   ├── Count dependencies
   ├── Check for critical patterns (auth, security, encryption)
   └── Assign classification

2. READ & MAP
   ├── Parse AST to extract structure
   ├── Identify classes and their methods
   ├── Identify standalone functions
   ├── Find global variables and constants
   └── Detect state mutations

3. DEPENDENCY CHECK
   ├── Internal imports (from project modules)
   ├── External imports (third-party)
   └── External calls (database, network, filesystem, messaging, ipc)

4. CONTEXT ANALYSIS
   ├── Where are exported symbols used?
   ├── What modules import this file?
   └── What message types flow through here?

5. RUNTIME VERIFICATION (if Critical/High-Complexity)
   ├── Use log analysis to trace actual behavior
   ├── Verify documented flow matches actual flow
   └── Note any discrepancies

6. DOCUMENTATION
   ├── Update analysis_progress.json
   ├── Generate module report section
   └── Cross-reference with CONTEXT.md

Runtime Verification Integration

For runtime verification of critical/high-complexity files, use your project's log aggregation system:

Trace actual behavior through components using correlation IDs
Verify that documented flows match actual runtime behavior
Use distributed tracing or structured logs to follow request paths

The goal is to confirm that code paths match documented behavior through runtime evidence.

Output Interpretation

JSON Output Structure

{
  "file": "src/utils/circuit_breaker.py",
  "classification": "critical",
  "metrics": {
    "lines_of_code": 245,
    "num_classes": 2,
    "num_functions": 8,
    "num_dependencies": 12
  },
  "structure": {
    "classes": [...],
    "functions": [...],
    "constants": [...]
  },
  "dependencies": {
    "internal": [...],
    "external": [...],
    "external_calls": [...]
  },
  "usages": [...],
  "verification_required": true
}

Markdown Output Format

The markdown output follows the template in templates/analysis_report.md and produces sections suitable for inclusion in phase deliverable documents.

Best Practices

Source Code Analysis (Phases 1-7)

Start with Phase 1: Foundation modules inform understanding of everything else
Track Progress: Always use --update-progress when completing analysis
Verify Critical Files: Never skip runtime verification for critical/high-complexity
Cross-Reference: After analysis, update CONTEXT.md links
Document Drift: Note any discrepancies between existing docs and actual code

Documentation Maintenance (Phase 8)

Run scan first: Always start with doc_review.py scan to understand current state
Fix links before content: Broken links indicate structural issues to address first
Verify against code: Never update documentation without verifying against actual source
Update indexes last: Navigation indexes should reflect final state after all changes
Generate health report: Always produce doc_health_report.json as evidence of completion

Documentation Maintenance Workflow

When invoking Phase 8 documentation maintenance, follow this sequence:

1. PLANNING
   ├── Run: doc_review.py scan --path docs/
   ├── Review health report
   ├── Identify priority fixes (broken links, obsolete files)
   └── Create todo list with specific actions

2. EXECUTION (in batches)
   ├── Batch 1: Fix broken links
   │   └── Run: doc_review.py validate-links --fix
   ├── Batch 2: Verify critical docs against source
   │   └── Run: doc_review.py verify --doc <file> --source <code>
   ├── Batch 3: Delete obsolete files
   │   └── Manual review + deletion
   ├── Batch 4: Update navigation indexes
   │   └── Run: doc_review.py update-indexes
   └── Batch 5: Update timestamps
       └── Set last_updated on verified files

3. VERIFICATION
   ├── Run: doc_review.py scan (confirm improvements)
   ├── Run: doc_review.py validate-links (confirm zero broken)
   └── Generate final doc_health_report.json

Resources

Scripts: scripts/ - Python analysis tools
- analyze_file.py - Source code analysis (Phases 1-7)
- check_progress.py - Progress tracking
- doc_review.py - Documentation maintenance (Phase 8)
- comment_rewriter.py - Comment analysis engine (antirez standards)
- rewrite_comments.py - Comment quality CLI tool
Templates: templates/ - Output templates
- analysis_report.md - Module-level report template
- semantic_analysis.md - AI-powered per-file analysis template
References: references/ - Analysis methodology docs
- DEEP_DIVE_PLAN.md - Master analysis plan with all phase definitions
- ANTIREZ_COMMENTING_STANDARDS.md - Complete antirez comment taxonomy
- AI_ANALYSIS_METHODOLOGY.md - AI semantic analysis methodology
- SEMANTIC_PATTERNS.md - Pattern recognition guide for Claude
analysis_progress.json: Progress tracking state
doc_health_report.json: Documentation health metrics (generated)
comment_health.md: Comment quality report (generated)

deep-dive-analysis

Deep Dive Analysis Skill

Overview

Capabilities

Prerequisites

CRITICAL PRINCIPLE: ABSOLUTE SOURCE OF TRUTH

IMPORTANT LIMITATION: Verification is Multi-Layer

The Iron Law of Documentation

Mandatory Rules (VIOLATION = FAILURE)

Verification Requirements

Documentation Verification Status

CRITICAL PRINCIPLE: NO HISTORICAL DEPTH

Available Commands

1. Analyze Single File

2. Check Progress

3. Find Usages

4. Generate Phase Report

Phase 8: Documentation Review Commands

5. Scan Documentation Health

6. Validate Links

7. Verify Against Source Code

8. Update Navigation Indexes

9. Full Documentation Maintenance

Comment Quality Commands (Antirez Standards)

10. Analyze Comment Quality

11. Scan Directory for Comment Issues

12. Generate Comment Health Report

13. Rewrite Comments

14. View Standards Reference

Comment Type Classification

Comment Quality Workflow

File Classification Criteria

AI-Powered Semantic Analysis

The Semantic Analysis Mandate

Five Layers of Understanding

Semantic Analysis Questions

Pattern Recognition

Red Flags to Identify

Semantic Analysis Template

AI Analysis Workflow

Reference Documents

Analysis Loop Workflow

Runtime Verification Integration

Output Interpretation

JSON Output Structure

Markdown Output Format

Best Practices

Source Code Analysis (Phases 1-7)

Documentation Maintenance (Phase 8)

Documentation Maintenance Workflow

Resources

Similar Skills