Claude-GPT Workflow Skills

中文 | English

A collection of Claude Code skills for Claude + GPT automatic collaboration workflows.

Skills

1. codex

Delegates coding tasks to Codex CLI for execution. CodeX is a cost-effective, strong coder — great for batch refactoring, code generation, multi-file changes, and multi-turn implementation tasks.

Based on oil-oil/codex.

Usage:

~/.claude/skills/codex/scripts/ask_codex.sh "Your request"

2. plan-review

Reviews a technical plan via Codex and iteratively refines it. Uses adversarial review to improve plan quality before implementation.

Trigger: /plan-review <plan-file-path>

3. plan-execute

Executes a finalized plan by delegating coding to Codex. Claude orchestrates, Codex codes, Claude reviews, Codex fixes — iterating until quality passes.

Trigger: /plan-execute <plan-file-path>

Note

After Claude enters plan mode, plan files are stored under ~/.claude by default. You can change the plan storage directory in ~/.claude/settings.json; for example, the config below stores plans in the current project's ./plans directory:

{
  "env": {},
  "plansDirectory": "./plans"
}

Installation

Option 1: npx skills add (Recommended)

npx skills add longranger2/claude-gpt-workflow

Option 2: Manual installation

Copy skills to your Claude Code skills directory:

cp -r plan-review/ ~/.claude/skills/
cp -r plan-execute/ ~/.claude/skills/
cp -r codex/ ~/.claude/skills/

Workflow

%%{init: {'theme': 'base', 'themeVariables': { 'clusterBkg': 'transparent', 'background': 'transparent', 'primaryColor': '#FFF7E6', 'primaryTextColor': '#5A3A00', 'primaryBorderColor': '#F59E0B', 'lineColor': '#6366F1', 'secondaryColor': '#E8F5FF', 'tertiaryColor': '#F3E8FF'}}}%%
flowchart TB
  %% ---------- Styles ----------
  classDef user fill:#E8F5FF,stroke:#1B6EF3,stroke-width:2px,color:#0B2A5B,stroke-dasharray: 0;
  classDef claude fill:#FFF7E6,stroke:#F59E0B,stroke-width:2px,color:#5A3A00,stroke-dasharray: 0;
  classDef codex fill:#F3E8FF,stroke:#7C3AED,stroke-width:2px,color:#2E1065,stroke-dasharray: 0;
  classDef decision fill:#FFFFFF,stroke:#6366F1,stroke-width:2px,color:#111827,stroke-dasharray: 5 5;

  %% ---------- Phase 1: Plan Review ----------
  subgraph P1["Phase 1: Plan Review (Adversarial Iteration)"]
    direction TB
    U1["User: Submit Plan"]:::user
    C1["Claude: Delegate to Codex for Critical Review"]:::claude
    X1["Codex: Output Review (10+ issues, Critical/High/Medium/Low)"]:::codex
    C2["Claude: Evaluate & Refine Plan"]:::claude
    D1{"Status?"}:::decision

    U1 --> C1 --> X1 --> C2 --> D1
    D1 -- "NEEDS_REVISION (Loop A)" --> C1
  end

  %% ---------- Phase 2: Plan Execute ----------
  subgraph P2["Phase 2: Plan Execute (Orchestrated Implementation)"]
    direction TB
    C3["Claude: Dispatch batch to Codex for implementation"]:::claude
    X2["Codex: Implement code changes (Executor)"]:::codex
    C4["Claude: Read-only Code Review (Alignment / Quality / Build)"]:::claude
    C5["Claude: Write Review (issue list)"]:::claude
    D2{"Verdict?"}:::decision
    D3{"More steps remaining?"}:::decision
    Done([Complete]):::decision

    C3 --> X2 --> C4 --> C5 --> D2
    D2 -- "NEEDS_FIX (Loop B)" --> X2
    D2 -- "APPROVED" --> D3
    D3 -- "YES (Loop C: next batch)" --> C3
    D3 -- "NO" --> Done
  end

  %% ---------- Phase Transition ----------
  D1 -- "APPROVED / MOSTLY_GOOD" --> C3

Loading

Core Concepts

Concept	Description
Adversarial	Codex acts as a "nitpicker", not an assistant — its job is to find flaws
Iterative	Not one-shot; multiple rounds of back-and-forth until quality gates pass
Role Separation	User defines what, Claude orchestrates how, Codex executes
Feedback Loops	Review → Fix → Re-review cycles (Loops A, B, C)

The Three Loops

• Loop A (Plan Refinement): Review finds issues → Refine plan → Re-review → ... → APPROVED
• Loop B (Code Fixing): Code review finds bugs → Codex fixes → Re-review → ... → APPROVED
• Loop C (Batch Processing): Complete current batch → Next batch → ... → All done

License

MIT