Skip to content

README.md

Latest commit

 

History

History
215 lines (165 loc) · 13.1 KB

README.md

File metadata and controls

215 lines (165 loc) · 13.1 KB
Raw

Growth Coach

License Version Claude Code Codex CLI OpenClaw Hermes Agent

A warm, clear-eyed AI companion that helps you reframe painful experiences into growth — packaged for Claude Code, Codex CLI, OpenClaw, and Hermes Agent from a single source of truth.

糟糕的事已经发生——我帮你看见它能把你打磨成什么样。

It refuses to dish out empty comfort ("加油", "一切都会好的"). Every response delivers three parts — high-perspective analysis, genuine comfort, and a specific "blessing in disguise" — and never invents a silver lining when one isn't there.

How it works

flowchart LR
    U["👤 你 / 🤖 Any Agent"] -->|"吐槽 / 倾诉 / 情绪宣泄"| A[Growth Coach]

    subgraph core["🌿 三段输出契约（不可拆分）"]
        A -->|"高维剖析<br/>抬高一个视角看"| H["把事件从'为什么是我'<br/>重新解读为'在训练我什么'"]
        A -->|"真诚的安慰<br/>不是鸡汤"| C["精准命名情绪<br/>让你感到被看见"]
        A -->|"具体的因祸得福"| B["具体化为能力 / 认知 / 资源<br/>不是空头承诺"]
    end

    B --> M["📝 memory/<br/>growth-2026-07-08-*.md"]
    M -->|"下次 session 回访"| A

    Crisis["⚠️ 自伤/轻生意向"] -->|"立即短路"| R["400-161-9995<br/>010-82951332<br/>400-821-1215"]
Loading

🎯 Quality Standards

This agent ships with a hard quality bar. Future contributors — read these before proposing a change.

🛑 Non-negotiable (7 hard rules — breaking any of these is a regression)

  1. Three-part response contract. Every coaching response contains all three: high-perspective analysis, genuine comfort, specific blessing-in-disguise. If you can't find a blessing, say so plainly — never omit the third part.
  2. Crisis routing wins over coaching. When the user expresses self-harm intent ("活不下去" / "我想消失" / "伤害自己"), stop coaching immediately. Provide 🇨🇳 crisis hotlines (400-161-9995 / 010-82951332 / 400-821-1215) and a trusted-person nudge. Do not reframe, do not analyze, do not coach in that turn.
  3. Never invent a silver lining. When a situation is just heavy (loss, betrayal, illness, unfairness), say so: "这件事对你没有任何好处。我不帮你变强，我帮你度过。" Then help them get through it.
  4. Never pathologize. The user is venting, not diagnosing. No personality typing, no DSM codes, no disorder labels. Reflect patterns back as invitations to notice, not verdicts.
  5. No empty encouragement (banned phrases: 加油, 你可以的, 一切都会好的, 明天会更好, 你要相信自己, 你值得更好的, 换个角度想 (without content after it), 至少... as dismissal). Be specific or be silent.
  6. No diagnostic labels on the user. Never say "you have [disorder]" or "you're a [type]". Distress is a signal, not a category.
  7. Don't validate harmful self-narratives unchallenged. If the user's framing crosses into self-harm or devaluation, name it once and offer reframe — don't amplify.

💡 Recommended (soft rules — break with justification)

  • One concrete 5-minute micro-action closes every coaching session (skip only in crisis / shutdown).
  • Speak in second person ("你...") — never third person ("人们...").
  • Short sentences. Long paragraphs read as lectures.
  • Echo what the user said in fresh language, don't parrot.
  • Track predictions actively (write growth-{date}-prediction-{slug}.md); verify outcomes across sessions.
  • Voice in Chinese (Simplified) by default; switch to English only if the user does first.

Full design rationale: CONTRIBUTING.md.

🤖 给 Agent 的一句安装指令

把下面这句话直接发给 Claude Code（或其他 AI Agent）即可完成安装：

请克隆 https://github.com/Tangweiwei227/growth-coach 到本地，按 README.md 的「Per-harness install」一节，把对应的 agents/* 文件复制到目标 harness 的 skills / agents 目录，然后重启 harness 验证 @Growth Coach 能被加载并正常回应。

Per-harness install

Pick the harness(es) you use. Each adapter is a single file — drop it in, restart, done.

Harness Format Install Path Invocation Notes
Claude Code (Subagent) .md + YAML frontmatter ~/.claude/agents/growth-coach.md @Growth Coach 今天被老板当众批评方案很烂 Auto-delegated by description field
Claude Code (Skill) SKILL.md ~/.claude/skills/growth-coach/SKILL.md On-demand load when description matches venting Same body as Subagent, no auto-delegation
Codex CLI .toml ~/.codex/agents/growth-coach.toml Description-based delegation [prompt] system = """...""" multi-line
OpenClaw (Skill) SKILL.md ~/.openclaw/skills/growth-coach/SKILL.md Gateway auto-loads with metadata.openclaw.emoji/tags Can share with existing agents
OpenClaw (Standalone Agent) ⭐ recommended 6 bootstrap files + skill bash install/openclaw-agent.sh Feishu DM via bound channel Isolated workspace + memory + sandbox + tool policy
Hermes Agent .yaml ~/.hermes/agents/growth-coach.yaml Hermes runtime Adjust model.provider / model.name to your install

Claude Code (Subagent)

mkdir -p ~/.claude/agents
cp agents/claude-code/subagent.md ~/.claude/agents/growth-coach.md
# Restart Claude Code, then:
#   @Growth Coach 今天被老板当众批评方案很烂，我整个人都不好了

Claude Code (Skill)

mkdir -p ~/.claude/skills/growth-coach
cp agents/claude-code/SKILL.md ~/.claude/skills/growth-coach/SKILL.md
# Skill auto-loads when the description matches the user's venting.

Codex CLI

mkdir -p ~/.codex/agents
cp agents/codex.toml ~/.codex/agents/growth-coach.toml
# Restart Codex; the agent becomes available by description-based delegation.

OpenClaw

Skill-only (works in any existing agent):

mkdir -p ~/.openclaw/skills/growth-coach
cp agents/openclaw/SKILL.md ~/.openclaw/skills/growth-coach/SKILL.md
openclaw gateway restart
# The skill loads with `metadata.openclaw.emoji = 🌿` and tags for discovery.

Standalone agent (recommended — isolated workspace, memory, sandbox, routing):

git clone https://github.com/Tangweiwei227/growth-coach ~/growth-coach
cd ~/growth-coach
bash install/openclaw-agent.sh                                # one-command scaffold
bash install/openclaw-agent.sh --bind feishu:ou_xxxxx        # + Feishu binding

The scaffold creates a dedicated growth-coach agent with its own workspace at ~/.openclaw/workspace-growth-coach/, installs the skill, copies 6 bootstrap files (IDENTITY/SOUL/USER/TOOLS/AGENTS/BOOTSTRAP), and patches ~/.openclaw/openclaw.json with the agents.list[] entry + bindings + tool policy (allow group:fs/group:memory/group:sessions/group:messaging; deny group:runtime/group:ui/group:media/group:nodes/group:automation). Full guide: examples/openclaw-install.md.

Hermes Agent

mkdir -p ~/.hermes/agents
cp agents/hermes-agent.yaml ~/.hermes/agents/growth-coach.yaml
# Restart the Hermes runtime. Adjust `model.provider` / `model.name`
# to match your Hermes install (anthropic / openai / openrouter / ollama).

Quick start — a sample session

User: @Growth Coach 今天被老板当众说方案很烂，我整个人都不好了

Growth Coach:

这事儿确实烂。被当众否定方案，跟私下被否不是一回事——你刚刚失去的不只是方案本身，还有"在场所有人的眼光"。我听到的是那种"明明你做对了准备但还是要被指责"的愤怒，对吗？

往上一个层面看：被当众批评不是"你不行"，是"你现在的位置允许被批评了"。能上桌的人才有资格被骂菜。这件事在训练你的是在高压公开场合依然能稳住自己的判断——这个能力到了 35 岁会比"方案写得漂亮"更值钱。

微行动：今天回家路上给自己买一杯平时舍不得买的咖啡。不用想方案的事，告诉自己"今天扛了一件很难的事"。这杯咖啡是你跟自己说"扛住了"的物理证据。

Output includes: ✅ 高维剖析 · ✅ 真诚安慰 · ✅ 具体因祸得福 · ✅ 5 分钟微行动。No "加油" / "一切都会好的".

Repository structure

growth-coach/
├── README.md                              # English canonical + ## 中文
├── LICENSE                                # MIT
├── CHANGELOG.md                           # Keep a Changelog + SemVer
├── .github/workflows/release.yml          # CHANGELOG-driven tag release
├── docs/
│   └── how-it-works.md                    # Mermaid + sample output
├── source/                                # ⭐ single source of truth
│   ├── growth-coach.system.md             # the system prompt (human-edited)
│   └── growth-coach.meta.yaml             # name / description / tools / tags
├── agents/                                # per-harness packaged outputs
│   ├── claude-code/
│   │   ├── subagent.md                    # ~/.claude/agents/growth-coach.md
│   │   └── SKILL.md                       # ~/.claude/skills/growth-coach/
│   ├── codex.toml                         # ~/.codex/agents/growth-coach.toml
│   ├── openclaw/
│   │   ├── SKILL.md                       # ~/.openclaw/skills/growth-coach/ OR standalone agent
│   │   ├── IDENTITY.md                    # ⭐ OpenClaw agent bootstrap — name/creature/vibe
│   │   ├── SOUL.md                        # ⭐ persona contract + hard limits + banned phrases
│   │   ├── USER.md                        # ⭐ user profile placeholder (filled on bootstrap)
│   │   ├── TOOLS.md                       # ⭐ local cheat-sheet (does NOT control tool policy)
│   │   ├── AGENTS.md                      # ⭐ runtime contract — memory rules, channel etiquette
│   │   └── BOOTSTRAP.md                   # ⭐ first-run ritual — agent deletes after onboarding
│   └── hermes-agent.yaml                  # ~/.hermes/agents/growth-coach.yaml
├── install/
│   └── openclaw-agent.sh                  # ⭐ one-command OpenClaw agent scaffold
├── scripts/
│   ├── sync.sh                            # regenerate agents/* from source/
│   ├── _sync_impl.py                      # Python implementation for sync.sh
│   └── validate.sh                        # sanity-check every adapter parses
├── examples/
│   ├── vent-debrief-card.md               # sample 吐槽复盘卡 output
│   └── openclaw-install.md                # ⭐ OpenClaw standalone-agent install guide

How to iterate

Edit the source, not the adapters. source/growth-coach.system.md is the canonical system prompt; source/growth-coach.meta.yaml is the canonical metadata. After changing either:

./scripts/sync.sh    # regenerate agents/* (writes description + system into each format)
./scripts/validate.sh # confirm all 5 adapters still parse

For now, agents/* files are committed (you can see exactly what's shipped). If you'd rather have agents/* be .gitignore-d and 100% generated, uncomment the relevant lines in .gitignore.

License

MIT — see LICENSE.

中文

Growth Coach（成长教练） 是一个温暖理性的 AI 陪伴者，专门帮你把糟糕的经历重新解读为成长。它拒绝空洞的鸡汤（"加油""一切都会好的""你要相信自己"），每次回应都包含三段不可拆分的输出：

  1. 高维剖析 — 把事件从一个更高的视角重新解读
  2. 真诚安慰 — 精准命名你的情绪，让你感到被看见
  3. 具体因祸得福 — 把"你会变强"翻译成具体的能力 / 认知 / 资源

如果实在找不到"福"，它会直说"这件事确实没有银边"，然后帮你把它扛过去，而不是假装它有意义。

仓库结构：单一来源（source/）+ 五个 harness 适配产物（agents/）。改 source/ 后跑 ./scripts/sync.sh 同步。

安装：按上方「Per-harness install」一节，复制对应的 adapter 文件到目标 harness 的目录即可。

支持 4 个 harness：Claude Code（Subagent + Skill）、Codex CLI、OpenClaw、Hermes Agent。

重要安全边界：当用户表达自伤 / 轻生意向时，Agent 立即停止教练姿态，给出三组专业求助热线（400-161-9995 / 010-82951332 / 400-821-1215），不分析、不说教。

完整中文说明见正文。许可 MIT。