ceo-audit-2-models

name

ceo-audit-2-models

description

Strategic project audit with a dual-model C-level council: Claude Opus + GPT-5.4 run as independent experts in parallel, then cross-model synthesis surfaces agreements, contradictions, and blind spots neither model catches alone. Use when you need a rigorous multi-perspective review before a major decision, launch, or architecture change. Automatically installs the OpenAI Codex plugin if not present. Triggers on: /ceo-audit-2-models, "audit the project", "get expert opinions", "multi-model review", "council review", "ceo council", "strategic review", "second opinion on the project".

CEO Audit — Dual-Model C-Level Council

Launch parallel sub-agents as isolated C-level experts. Every expert receives identical project data. No coordination between agents — isolation produces genuine diversity of opinion. Then run the same roles on both Claude Opus and GPT-5.4 simultaneously and synthesize where they agree, disagree, and diverge.

Cross-model mode is the default and recommended mode: overlapping findings get high confidence, model-unique findings reveal blind spots.

---

Phase 0: Check & Install Codex Plugin

GPT-5.4 experts run through the OpenAI Codex plugin. Check availability before launching any experts.

CODEX=$(find ~/.claude/plugins/cache -name "codex-companion.mjs" 2>/dev/null | sort -V | tail -1)
echo "${CODEX:-NOT_FOUND}"

If NOT_FOUND: install the plugin now:

/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup

Then re-run the find command to get the path. Store it in $CODEX for all subsequent GPT calls.

If found: proceed immediately. No setup needed.

---

Phase 1: Scan Project Context

Before suggesting experts, understand the project:

1. Read CLAUDE.md (or README.md if absent)
2. Scan .claude/rules/ for domain context
3. Glance at top-level file structure

Based on findings, generate 4-6 expert roles tailored to this specific project. Roles must reflect the project's actual domain, challenges, and current stage — not generic defaults.

Role Examples by Project Type

Don't copy these — generate fresh roles based on actual context:

Project Type	Typical Roles
SaaS	Head of Engineering, Head of Product, Head of Growth, CFO, UX Researcher
Open Source	Community Manager, Technical Architect, DevRel, Security Advisor
Content / Media	Content Strategist, Audience Analyst, Monetization Expert, Distribution Expert
EdTech	CMO, CFO, CPO, COO, Growth Advisor
E-commerce	Head of Supply Chain, Marketing Director, CTO, Customer Experience Lead
Agency / Consulting	Sales Director, Delivery Lead, Talent Manager, CFO

---

Phase 2: Assemble the Council

Always ask the user before proceeding. Do not pick roles yourself.

2a. Choose Model Mode

Ask via AskUserQuestion (single-select):

Mode	Description
Cross-model (recommended)	Each role runs on Claude Opus AND GPT-5.4; cross-model synthesis highlights blind spots
Claude only	All experts on Claude Opus — faster, no Codex required
GPT only	All experts on GPT-5.4 via Codex

2b. Choose Roles

Use AskUserQuestion with multiSelect: true. Show the 4-6 role options you generated in Phase 1 with short descriptions of their focus. The user can always write custom roles via "Other".

Minimum 2 experts. If the user picks only 1, suggest adding one more — productive disagreement requires at least two perspectives.

---

Phase 3: Collect Project Data

Gather project state to feed all experts. Prepare this block once — it must be identical for every expert regardless of model.

Read:

• Key metrics and data files identified during Phase 1
• Strategy and planning documents
• Recent decisions (e.g., git log --oneline -10)
• Previous council analyses (if any exist in docs/)

Skip: GitHub stars, clone counts, contributor lists — vanity metrics add noise without strategic signal.

Store the result as DATA_BLOCK. Every prompt in Phase 4 uses the same DATA_BLOCK verbatim.

---

Phase 4: Generate Expert Prompts

Claude prompts (natural language)

You are the [ROLE] for [PROJECT NAME]. Analyze the data below from a [DOMAIN] perspective.

Focus on:
- [3-6 specific focus areas relevant to role and project]

Data:
[DATA_BLOCK — identical for all experts]

[Role-specific personality instruction: "show the math", "be the contrarian",
"prioritize by effort/impact ratio", etc.]

Respond in the same language as the data provided.

GPT-5.4 prompts (XML-structured)

Only in Cross-model and GPT-only modes.

For each role, generate the XML prompt using gpt-prompt-template.md. Write each to a temp file:

cat > /tmp/ceo-audit-gpt-{role-slug}.txt << 'PROMPT'
[XML prompt content — see gpt-prompt-template.md for structure]
PROMPT

Rules:

• Same DATA_BLOCK, same focus areas, same personality instruction as the Claude version
• Only the format differs (natural language vs XML-structured)
• Use --prompt-file flag to avoid shell quoting issues with long prompts

---

Phase 5: Execute — ALL Experts in ONE Message

Launch everything simultaneously. One message = true parallelism.

Claude-only mode

Task(subagent_type: "general-purpose", model: "opus", prompt: "<CFO prompt>", description: "CFO analysis")
Task(subagent_type: "general-purpose", model: "opus", prompt: "<CTO prompt>", description: "CTO analysis")

Cross-model mode

Write GPT prompt files first (via Bash), then launch all experts in a single message:

# Claude experts
Task(subagent_type: "general-purpose", model: "opus", prompt: "<CFO Claude prompt>", description: "CFO (Claude Opus)")
Task(subagent_type: "general-purpose", model: "opus", prompt: "<CTO Claude prompt>", description: "CTO (Claude Opus)")

# GPT-5.4 experts
Bash(command: 'node "$CODEX" task --json --prompt-file /tmp/ceo-audit-gpt-cfo.txt',
     description: "CFO (GPT-5.4)", timeout: 300000)
Bash(command: 'node "$CODEX" task --json --prompt-file /tmp/ceo-audit-gpt-cto.txt',
     description: "CTO (GPT-5.4)", timeout: 300000)

GPT-only mode

Same as above but only the Bash calls (no Task tool calls).

Fallback

If Codex is unavailable (auth expired, CLI missing, network error):

⚠️ GPT-5.4 unavailable: [reason]
Switching to Claude-only mode.

Wait for all experts to return results before proceeding to Phase 6.

---

Phase 6: Synthesize

The synthesis is the entire value of the council. Never skip it.

Claude-only synthesis

# Council Session: [DATE]

## Council Members
[List of selected experts and their focus]

## Context
[Current state snapshot — brief]

## [Expert 1]
[Key findings and recommendations]

## [Expert 2]
[Key findings and recommendations]

## Consensus (all agree)
1. ...

## Disagreements
| Expert | Position | Argument |
|--------|----------|----------|

## Decisions
_To be filled after discussion._

Cross-model synthesis

GPT responses arrive as JSON — extract the rawOutput field for each expert.

For each role, compare Claude vs GPT findings side by side:

# Cross-Model Council Session: [DATE]

## Council Members
| Role | Claude Opus | GPT-5.4 |
|------|-------------|---------|
| CFO | ✅ | ✅ |
| CTO | ✅ | ✅ |

## Context
[Current state snapshot — brief]

## Per-Role Comparison

### [ROLE] Analysis

**Claude Opus:**
[Key findings — 3-5 bullets]

**GPT-5.4:**
[Key findings — 3-5 bullets]

**Alignment:** 🟢 Agree / 🟡 Partial / 🔴 Contradict
**Details:** [what matches, what diverges]

---

## Cross-Model Synthesis

### 🟢 Overlapping Findings (both models agree)
> High confidence — act on these first

1. [Finding] — both models converge
2. ...

### 🔵 Claude-Only Findings
> Potential GPT blind spots — verify before dismissing

1. [Finding] — Claude sees, GPT misses
2. ...

### 🟣 GPT-Only Findings
> Potential Claude blind spots — mandatory check

1. [Finding] — GPT sees, Claude misses
2. ...

### 🔴 Direct Contradictions
| Topic | Claude position | GPT position | Recommendation |
|-------|----------------|-------------|----------------|

## Model Agreement Score
- Full overlap: N findings
- Partial: N findings
- Contradictions: N findings
- **Agreement rate:** X%

## Top Recommendations (by confidence)
1. [Overlapping — highest confidence]
2. ...

## Decisions
_To be filled after discussion._

---

Phase 7: Save

Save to docs/council-[DATE].md (or a project-specific path if context suggests one).

---

Common Mistakes

Mistake	Fix
Picking roles without asking user	Always use AskUserQuestion first
Using Bash to launch Claude experts	Only use Task tool with subagent_type: "general-purpose"
Using Task tool for GPT experts	Use Bash + node "$CODEX" for GPT
Giving experts different data	Prepare ONE DATA_BLOCK, paste into all prompts verbatim
Different focus areas between models	Same role must have identical focus on both models
Gathering vanity metrics	Focus on strategy docs, actual metrics, recent decisions
Too many experts (6+ per model)	3-4 per model is optimal for signal-to-noise
Skipping synthesis	The synthesis IS the value
Skipping cross-model comparison	In cross-model mode, the comparison is the entire point
Not checking Codex availability	Always verify $CODEX path before launching GPT experts
Passing GPT prompt as CLI arg	Always use --prompt-file to avoid shell quoting issues
Hardcoding paths to codex-companion	Always discover dynamically via find ~/.claude/plugins/cache