English | 中文

fireworks-half-vibe

AI writes the scaffolding; you write the soul. A Claude Code / Codex skill that keeps your real programming skill alive in the age of vibe coding — without ever feeling like homework.

The Problem

Vibe coding ships fast — and quietly erodes the very skill that makes you good at it. Six months of "AI, write this for me" and the muscle is gone: you can review output, but you can no longer produce it.

fireworks-half-vibe is a coach layer that lives inside your normal AI coding session. The agent still does the heavy lifting — but at the moments that actually matter (the core algorithm, the state transition, the edge case), it hands you the pen. You write the pseudocode, it reviews like a pair-programming partner, and then your own logic ships as real code.

No drills. No quizzes. No report cards. Just your project, moving forward, with the steering wheel in your hands on the interesting stretches.

How It Works

You: "帮我写一个带重试的请求队列"
  → Agent frames the problem, marks 1–3 hand-write zones
  → You hand-write pseudocode for the core logic (chat fragments are fine)
  → Agent reviews like a colleague: "这个思路对——空队列进来那下会走到哪？"
  → Your pseudocode becomes real code, visibly yours
  → Agent silently logs your growth and tunes tomorrow's difficulty

Works in three modes: build (write code), decode (master a technical article or video by rebuilding its core mechanism), concept (turn any technical question into a tiny hands-on sketch).

Core Cases

Full annotated sessions live in examples/ — here is the flavor:

Build — you write the retry policy, AI writes everything else:

You:   等 2的n次方秒, 加个随机抖动避免惊群
Agent: 漂亮 —— 指数退避 + jitter 都对了。追问: n 一直涨, 等待时间会爆炸, 怎么办?
You:   封顶, 最多 30 秒
Agent: 对, cap。Zone 通过 —— 你的决策合成代码:
       min(2 ** attempt + random.uniform(0, 1), 30)

Decode — master a famous article by rebuilding its core, not re-reading it:

Agent: orchestrator-workers 的循环留给你: 拆解、分派、收尾, 三个空。
You:   (填了三个, 第四个卡住) 直接给我看完整的
Agent: (不给脸色, 完整示范 + 只要求一件事) 用自己的话讲一遍:
       "拆解权在 LLM 手里"和"在开发者手里"差在哪?
You:   在 LLM 手里是动态决定顺序并行异步; 在开发者手里是提前写死静态路径

Switch any time — four switches in one afternoon, zero ceremony:

14:00  (pure vibe, shopping cart module)
14:20  "只有库存扣减用 half-vibe"   → scoped, 超卖防护留给你
15:10  "切回 vibe"                 → 剩余部分 AI 直接写完
15:40  "支付回调这段我自己写"       → on-demand, AI 立即停笔交棒

Five Ways To Combine It With Vibe Coding

Switch any time, mid-session, with one sentence:

Designed For Chat Windows

Hand-writing pseudocode doesn't mean typing clean code into a chat box:

• Messy fragments — write "遍历队列 如果空 break", the agent normalizes it
• Zone file — the agent scaffolds a file, you edit it in your own editor
• Assembled — you supply only decisions (choices, blanks, ordering)
• Dictation — narrate in plain language, the agent formalizes

It Learns You

A local learner profile (.half-vibe/ in your project) tracks your level per domain, your recurring error patterns, and your preferred coaching style. Difficulty auto-adjusts — challenge stays exactly at the edge of your ability, where flow lives. Stuck is data, not failure: hints are free, skips are legitimate, and the profile records patterns, never demerits.

Installation

git clone https://github.com/yizhiyanhua-ai/fireworks-half-vibe.git
cd fireworks-half-vibe
./install.sh

Or copy SKILL.md + references/ into ~/.claude/skills/fireworks-half-vibe/
(or ~/.codex/skills/, .agents/skills/) manually.

Quick Start

进入 half-vibe 模式          # start coaching
教我写代码而不是替我写        # same thing
这段我自己写                 # claim a zone mid-vibe
切回 vibe                   # back to full-speed AI, instantly
我的编程水平                # view your learner profile

See examples/quick-start.md for a full sample session.

Repository Layout

├── SKILL.md                     # the skill (agents read this)
├── references/
│   ├── pedagogy.md              # coaching method, difficulty engine, input channels
│   └── profile-schema.md        # learner profile format
├── examples/quick-start.md      # annotated sample session
├── docs/index.html              # project website
└── install.sh                   # installer for Claude Code / Codex / .agents

License

MIT — see LICENSE.