Claude Opus Workflows

Structured multi-model workflows for AI debates, deep analysis, and complex reasoning. Uses Claude Opus via OpenAI SDK wrapper.

When to Use This vs opus-subagent

Situation	Use
Quick code review	opus-subagent
Debate with multiple perspectives	This skill (ai-debate)
Simple debugging	opus-subagent
Deep analysis with token budget	This skill (code-review/analyze)
Second opinion	Either
Multi-model synthesis	This skill

Quick Reference

Setting	Value
Base URL	http://localhost:8000/v1
API Key	my-super-secret-password-123
Default Model	claude-opus-4-5

Decision Tree

1. Subjective Decision / Trade-off? → AI Debate (debate)
2. Code Quality Check? → Code Review (review)
3. Understanding Complex Code? → Deep Analysis (analyze)
4. Multi-Step Reasoning? → Think Deep (think)
5. Complex Bug? → Debug (debug)

Workflow 1: AI Debate (Prompt Chaining)

Multi-model debate using prompt chaining for stronger performance. Each step gets Claude's full attention.

Why Chain Prompts for Debates?

1. Accuracy: Each perspective gets dedicated reasoning
2. Clarity: Cleaner handoffs between viewpoints
3. Traceability: Easy to identify weak arguments
4. Self-correction: Final synthesis can reference and refine earlier points

When to Use

• Architecture decisions
• Library/framework choices
• Trade-off analysis
• Any subjective decision

Quick Usage

cd .claude/skills/ai-opus-debate
uv run python3 -m scripts.debate --topic "JWT vs Session auth for mobile" --rounds 3

Chained Debate Process

Prompt 1: Security Advocate

You are a Security Advocate. Argue for the most secure option.
<topic>{{TOPIC}}</topic>
Output your argument in <security_perspective> tags.

Prompt 2: Pragmatic Engineer (receives previous output)

Review the security perspective, then argue for practicality.
<security_perspective>{{SECURITY_OUTPUT}}</security_perspective>
<topic>{{TOPIC}}</topic>
Output your argument in <engineering_perspective> tags.

Prompt 3: UX Advocate (receives all previous)

Review both perspectives, then argue for user experience.
<security_perspective>{{SECURITY_OUTPUT}}</security_perspective>
<engineering_perspective>{{ENGINEERING_OUTPUT}}</engineering_perspective>
Output in <ux_perspective> tags.

Prompt 4: Synthesis (receives all perspectives)

Synthesize all perspectives into a final recommendation.
<security_perspective>{{SECURITY_OUTPUT}}</security_perspective>
<engineering_perspective>{{ENGINEERING_OUTPUT}}</engineering_perspective>
<ux_perspective>{{UX_OUTPUT}}</ux_perspective>
Grade each argument (A-F), identify the strongest points, and provide a clear recommendation.

Self-Correction Chain (Optional)

Add a verification step after synthesis:

Prompt 5: Review

Review this recommendation for logical consistency and completeness.
<recommendation>{{SYNTHESIS}}</recommendation>
<original_topic>{{TOPIC}}</original_topic>
Identify any gaps, contradictions, or unconsidered trade-offs.

Prompt 6: Refine

Refine the recommendation based on this feedback.
<recommendation>{{SYNTHESIS}}</recommendation>
<feedback>{{REVIEW_OUTPUT}}</feedback>
Output the final, improved recommendation.

See workflows/ai-debate.md for full details.

Workflow 2: Code Review

Token-efficient structured code review.

Quick Usage

# Using Python script
cd .claude/skills/ai-opus-debate
uv run python3 -m scripts.check --mode precommit --input "$(git diff HEAD~1)"

Gathering Context with MCP Tools

Task	MCP Tool	Parameters
Semantic search	claude-context/search_code	query: "authentication middleware"
Database schema	postgres/list_tables	table_names: "users,sessions"
Symbol refs	plugin_serena_serena/find_referencing_symbols	symbol_name_path: "AuthMiddleware"

See workflows/code-review.md for full details.

Workflow 3: Deep Analysis

Comprehensive code analysis (architecture, security, performance).

cd .claude/skills/ai-opus-debate
uv run python3 -m scripts.analyze --files src/*.py --type security

MCP Tools Reference

Task	MCP Tool	Parameters
Semantic code search	claude-context/search_code	query: "your search"
Database schema	postgres/list_tables	table_names: "table1,table2"
Symbol references	plugin_serena_serena/find_referencing_symbols	symbol_name_path: "Symbol"
Get symbol body	plugin_serena_serena/find_symbol	name_path_pattern: "Symbol", include_body: true
Web research	exa/web_search_exa	query: "search query"

Tool Priority:

1. claude-context/search_code - Semantic discovery (preferred)
2. plugin_serena_serena/* - Symbol navigation (specific code)
3. postgres/* - Database context (only if DB involved)

Python Scripts

All scripts in scripts/ directory:

Script	Purpose
debate.py	Run AI debates
analyze.py	Deep code analysis
check.py	Pre-commit validation
think.py	Deep reasoning
debug.py	Systematic debugging
plan.py	Strategic planning
assist.py	Coding assistance

Running Scripts

cd .claude/skills/ai-opus-debate

# Install deps (once)
uv pip install openai

# Run scripts
uv run python3 -m scripts.debate --topic "Topic" --rounds 3
uv run python3 -m scripts.analyze --files src/*.py --type security
uv run python3 -m scripts.check --mode precommit --input "diff"

Configuration

export CLAUDE_API_BASE_URL="http://localhost:8000/v1"
export PROXY_API_KEY="my-super-secret-password-123"
export CLAUDE_DEFAULT_MODEL="claude-opus-4-5"

Files

• SKILL.md - This reference
• workflows/ - Detailed workflow guides
  • ai-debate.md - Multi-model debate
  • code-review.md - Structured review
  • deep-code-analysis.md - Analysis workflow
• references/ - API and tool docs
• scripts/ - Python CLI scripts