AI operations system for Claude Code. Three agents collaborate to run your project's ops — strategy, content, community, data tracking, and intel — while you steer.
English | 中文
I need to do operations for my projects, but I'm not good at it. So I built a system where AI actually does the operations — not just suggests, but executes.
Most AI tools help you write content. This one runs the whole loop: strategy, content creation, community management, data tracking, and intel scanning. You set the direction, the agents execute.
ccops keeps its own ops diary — what it did each day and what it learned. No screenshots or feature lists, just the running record of a system that actually runs. That's the real proof. Full diary →
6/14 — Made the diary the thing we lead with. It's a continuous, honest record written every day — stronger than any one-off screenshot. Moved it up so visitors see it first.
6/13-14 — Baked a launch mishap into the skills: shipped a promo post while the repo was still private, so the link 404'd — then nobody checked why it got zero engagement. Added two forced steps: a pre-publish front-door check, and a post-publish zero-engagement follow-up. Wrote GROWING.md. Started joining Hacker News discussions on open-source AI tools.
6/11 — Helped define Steam strategy for FlyingSword (free demo, ongoing updates, Next Fest). Specialist autonomously finished a 6-dimension deep dive on competitor AiToEarn (20k stars). Store page copy hit v3 — first two read like feature lists, studied Exo One and Race the Sun first, then got it right.
I used it to promote another open source project. Here's what happened:
Set direction. Chatted with the Conversation Director for 30 minutes — what the project is, which platforms to target. It asked a dozen questions, wrote up a strategy. I tweaked a few lines, confirmed.
Did research. It scanned 200+ top posts on the target subreddit, figured out which title formats work, what timing gets traction. Checked out the competition too — who's using similar tools, what the community thinks.
Wrote content. 12 drafts from different angles. Then three AIs simulated target platform users and scored each one — style match, credibility, viral potential. I got a ranked scorecard.
Picked drafts. I chose 3 from the 12, tweaked a few lines, ready to publish.
My total contribution: set the direction, pick the drafts. The rest was the system.
Three roles, one system:
┌─────────────────────────────────────────────────────┐
│ You (human) │
│ Set direction · Approve publishes │
└──────────┬──────────────────────────────┬───────────┘
│ conversation │ run.sh (auto)
▼ ▼
┌─────────────────┐ ┌──────────────────┐
│ Conversation │ │ Auto Director │
│ Director │ │ (unattended) │
│ │ │ │
│ • Strategy │ │ • Review results│
│ • Skill design │ │ • Assign tasks │
│ • Approve │ │ • Extract learnings
│ publishes │ │ • Write reports │
└─────────────────┘ └────────┬─────────┘
│ assigns tasks
▼
┌──────────────────┐
│ Specialist │
│ (per project) │
│ │
│ • Execute tasks │
│ • Create content│
│ • Collect data │
│ • Scan intel │
└──────────────────┘
Two parallel lines, not sequential phases: conversation and auto run side by side. Published content never goes out without human confirmation.
Same knowledge, two modes. Both read the same strategy and memory files. Judgments don't drift.
Agents don't share context windows. They communicate through structured files. Simple, debuggable, inspectable.
projects/your-project/
├── knowledge/
│ ├── strategy.md # Strategy (director writes, specialist reads)
│ └── insights.md # Learnings (director extracts, specialist reads)
├── state/
│ └── tasks.md # Tasks + receipts (single file)
├── content/ # Generated content (by platform)
├── metrics/ # Data tracking (tracker.json)
└── intel/ # Competitive intelligence
Every agent starts fresh each cycle — all state is in files, not conversation history.
Ships with zero domain skills. The Conversation Director builds them as needed, using skill-forge. Each project grows its own skillset.
| Built-in Skill | Purpose |
|---|---|
ops-setup |
Initialize: interview → search → confirm strategy → hand off |
skill-forge |
Create and improve skills (the only way new skills grow) |
Use it once, it learns once. You figure out what works, it becomes a reusable skill.
Prerequisites: Claude Code CLI + GitHub CLI installed and authenticated. opencli with browser skill recommended for platform operations.
# 1. Clone
git clone https://github.com/ailess-lab/claude-code-ops.git
cd claude-code-ops
# 2. Initialize your project
# In Claude Code, run /ops-setup — it walks you through setup
# 3. Run auto ops
bash run.shSee QUICKSTART.md for the full guide. Once you're running, GROWING.md covers how to make ccops better the more you use it.
- Currently used by one person (me).
- Requires basic command line and Claude Code knowledge.
- Not plug-and-play — you spend 30 minutes setting strategy first, then it works.
- Publishing has a safety lock. Nothing goes out without your say-so.
- Show, don't tell. Real results from real projects.
- Honest perspective. Built by someone who needs operations but isn't an operations expert.
- Plain language. No jargon, no buzzwords.
- Useful > perfect. Ship something that works, iterate from there.
我做项目还行，做运营是真的不行。
不是不想做——每次打开 Reddit 想发帖，脑子就空白。写推广文案比写代码难十倍。数据追踪、社区互动、竞品分析，知道该做，就是做不来。
所以造了这个。说功能没感觉，说个真事吧。
ccops 每天给自己写运营日记——干了什么、学到了什么。不靠截图、不列功能，就是把一个真在跑的系统的运行记录如实摆出来。这是最硬的证明。完整日记 →
6/14 — 把运营日记本身定为主推。它是每天在长、如实记录的连续内容，比任何一次性的截图都更有说服力。提到首页，让访客第一眼看到。
6/13-14 — 把一次发布事故的教训固化进了 skill：仓库还没设公开就发了带链接的推广帖，读者点开是 404，事后零互动也没查为什么。加了两道强制步骤——发布前过门面检查、发布后零互动必须倒查。写了 GROWING.md。开始在 Hacker News 参与开源 AI 工具讨论。
6/11 — 协助 FlyingSword 定了 Steam 策略：免费 Demo、持续更新、目标参加 Next Fest。专员自主跑完了竞品 AiToEarn（2 万星）的六维度调研。商店页文案写到第三版——前两版太像功能列表被毙，研究竞品商店页才改对。
用它推了另一个开源项目。真实经过：
定方向。跟对谈主管聊了半小时，说清楚项目是什么、想在哪些平台发。它问了十几个问题，写了份运营策略。我改了几条，拍板。
做调研。它自己扫了目标社区前 200 个帖子，总结什么标题能火、什么格式受欢迎。竞品也查了一遍——谁在用同类工具、社区里什么态度。
写文案。12 篇，不同角度。写完不自己说好——三个 AI 分别模拟目标平台用户，每篇按风格、可信度、传播力打分。我拿到的是一份评分表。
选稿。我从 12 篇里挑了 3 篇，改了几句，准备发三个平台。
整个过程我做了什么？定方向、选稿子。其余它干的。
三个 AI 角色各管一摊：
┌─────────────────────────────────────────────────┐
│ 你（人） │
│ 定方向 · 发布前确认 │
└──────┬────────────────────────────┬─────────────┘
│ 对谈 │ run.sh（自动）
▼ ▼
┌─────────────┐ ┌──────────────────┐
│ 对谈主管 │ │ 自动主管 │
│ │ │ （无人值守） │
│ • 定策略 │ │ • 复盘结果 │
│ • 沉淀skill │ │ • 派任务 │
│ • 确认发布 │ │ • 提炼经验 │
└─────────────┘ └───────┬──────────┘
│ 派任务
▼
┌──────────────────┐
│ 专员 │
│ （按项目） │
│ │
│ • 执行任务 │
│ • 产出内容 │
│ • 采集数据 │
└──────────────────┘
两条线并行，不是先后阶段：对谈和自动同时跑。对外发布永远要你确认，不会偷偷发。
所有状态在文件里，不在对话里。关掉终端下次打开，从文件恢复，不丢任何东西。bash run.sh 一行启动。
出厂不带任何领域 skill。用一次，学一次。你跑通了什么，它就会什么。
前置：Claude Code CLI + GitHub CLI 安装并登录。opencli 搭配浏览器 skill 推荐，用于平台操作。
# 1. 克隆
git clone https://github.com/ailess-lab/claude-code-ops.git
cd claude-code-ops
# 2. 初始化你的项目
# 在 Claude Code 里跑 /ops-setup — 它会引导你完成设置
# 3. 跑自动运营
bash run.sh完整指南见 QUICKSTART.md。跑起来后，GROWING.md 讲怎么让 ccops 越用越好。
- 目前就我在用，没有"数千用户"。
- 需要懂一点命令行和 Claude Code。
- 不是开箱即用——先花半小时定策略，然后它才开始干活。
- 对外发布有安全锁，不会偷偷发东西出去。
- 展示，不吹。 真实项目的真实过程。
- 诚实视角。 造它的人需要运营但不擅长运营。
- 说人话。 不堆术语和空话。
- 有用比完美重要。 先跑起来，再迭代。