🇨🇳 中文版

claude-config

Personal Claude Code configuration with automated development pipeline. Clone and run to restore on a new machine.

---

Quick Start

1. Install prerequisites

npm install -g @anthropic-ai/claude-code
npm install -g @openai/codex

2. Clone and setup

git clone https://github.com/gyy0592/claude-config.git
cd claude-config

# Claude Code: writes CLAUDE.md, rules, system_override, and patches shell rc
bash set_claude.sh && source "${ZDOTDIR:-$HOME}/.zshrc" 2>/dev/null || source ~/.bashrc

# Codex CLI: fill in your base_url and api_key, then run setup
cp codex_info.yaml.template codex_info.yaml
# Edit codex_info.yaml → fill the 'now' section
bash setup_codex.sh

3. Install plugins (inside Claude Code)

cp settings.json ~/.claude/settings.json

# Run these inside Claude Code (slash commands, not terminal)
/plugin marketplace add humania-org/humanize
/plugin marketplace add openai/codex-plugin-cc
/plugin install humanize@humania
/plugin install codex@openai-codex

4. Install custom skills

mkdir -p ~/.claude/skills
for s in skills/*/; do
  ln -sfn "$(pwd)/skills/$(basename "$s")" ~/.claude/skills/"$(basename "$s")"
done

See docs/skills.md for the full skill list and triggers.

---

Automated Development Workflow

The core feature: a Humanize pipeline where Claude codes and Codex reviews in a closed loop.

draft.md  →  gen-plan  →  plan.md  →  RLCR loop  →  done

Step 1: Write a Draft

Create draft.md describing what and why. Include:

• Goal — what the feature/fix should accomplish
• Known Facts / Constraints — e.g. "GPU memory < 24GB", "follow experiment-run output conventions", "Python 3.10+"
• Rough acceptance criteria (gen-plan will formalize them)

/gen-draft

Step 2: Generate a Plan

/humanize:gen-plan --input draft.md --output docs/plan.md

Step 3: Review the Plan

Do not skip this. Read the plan. Verify acceptance criteria, task breakdown, path boundaries. If you have feedback, annotate with CMT: ... ENDCMT and refine:

/humanize:refine-plan --input docs/plan.md

Repeat until the plan is correct. The RLCR loop is an amplifier — a wrong plan executed flawlessly is still wrong.

Step 4: Run the RLCR Loop

/humanize:start-rlcr-loop docs/plan.md --codex-model gpt-5.3-codex --max 5

Flag	Purpose
--codex-model	Codex model for reviews (e.g. gpt-5.3-codex)
--max N	Max iterations before auto-stop
--yolo	Full automation: skip quiz + Claude answers Codex questions
--skip-quiz	Skip the plan understanding quiz only
--agent-teams	Parallel development with Agent Teams
--skip-impl	Skip to code review (for reviewing existing changes)

How the loop works: Claude implements → writes summary → Codex reviews → feedback loop until COMPLETE → codex review checks code quality → fix issues → done.

One-Shot (Plan + Loop)

/humanize:gen-plan --input draft.md --output docs/plan.md --auto-start-rlcr-if-converged

Monitor / Cancel

# Setup (one-time)
source ~/.claude/plugins/cache/humania/humanize/*/scripts/humanize.sh

# Monitor in another terminal
humanize monitor rlcr

# Cancel
/humanize:cancel-rlcr-loop

---

Sync

cp ~/.claude/settings.json settings.json
for s in skills/*/; do
  name=$(basename "$s")
  [ -d ~/.claude/skills/"$name" ] && rsync -a --delete ~/.claude/skills/"$name"/ skills/"$name"/
done
git add -A && git commit -m "sync" && git push

---

Credits

Ten Commandments for AI-Assisted Coding (EN | ZH) — adapted from Humanize by Dr. Sihao Liu, with personal extensions.