Pre-Vibe
Pre-Vibe turns a rough first request into three structured documents before real work begins. It reads your project safely, asks only what it needs to know, then hands Claude Code a compact execution contract — so every session starts with clear scope, project-aware guidance, and a focused prompt.
Designed for new Claude Code users, junior builders, and early product sessions where the first request is short, vague, or missing execution detail.
One-Click Install
In Claude Code, type:
帮我安装一下这个 plugin：KimFischer99/CC-Pre-Vibe
Or install manually:
# 1. Add the marketplace (one-time)
claude plugin marketplace add https://raw.githubusercontent.com/KimFischer99/CC-Pre-Vibe/main/.claude-plugin/marketplace.json
# 2. Enable auto-update
claude plugin marketplace auto-update --enable pre-vibe
# 3. Install the plugin
claude plugin install pre-vibe@pre-vibe
Restart Claude Code, type /pre-vibe, and tell it your rough idea.
Quick Start (3 Minutes)
| Step | What happens |
|---|
1. Type /pre-vibe in Claude Code | Plugin activates |
| 2. Say your rough idea | e.g. "Help me turn this idea into a build plan" |
| 3. Answer 2-5 questions | Claude Code opens native question dialogs |
| 4. Review the generated docs | Three files appear in your project |
| 5. Approve the handoff | /clear → inject FIRST_PROMPT.md → real work begins |
What Pre-Vibe Builds
Pre-Vibe writes three task-specific documents that together constrain Claude Code's context and improve implementation precision:
| Document | What it does | Why it matters |
|---|
PRE_VIBE_SPEC.md | Beginner-friendly project handbook — goals, scope, project language, evidence, risks, acceptance criteria | Aligns Claude Code with your domain terminology and success criteria before a single line of code |
CLAUDE.md (or PROJECT_CLAUDE.md) | Agent-facing execution rules — constraints, file pointers, operation boundaries, verification requirements | Loaded automatically by Claude Code at session start; prevents scope creep and guides every response |
FIRST_PROMPT.md | Compact execution contract with Completion Contract, stop/ask conditions, and verification gates | The only prompt after handoff — keeps Claude Code focused on the deliverable, not the setup |
PROJECT_INDEX.md is also available at architect effort level for high-uncertainty work.
These three documents work together to constrain large-model context, reduce hallucination, and keep Claude Code on track — especially important for non-English prompts, complex projects, or sessions with limited token budgets.
Effort Levels
| Level | Questions | Best for |
|---|
| mini | ≤ 3 | Small changes, quick fixes, simple research |
| default | ≤ 5 | Normal coding or research tasks |
| architect | ≤ 10 + PROJECT_INDEX.md | New products, refactors, high-uncertainty work |
Adjust anytime through the plugin settings:
- Default effort:
auto, mini, default, architect
- Session effort override
- Architect-only project index toggle
- Claude Code environment inspection toggle
How It Works
- Detects existing context,
CLAUDE.md files, and git state
- Builds a safe project index (skips secrets,
node_modules, etc.)
- Resolves effort level and document plan
- Opens native question UI for unresolved decisions
- Writes customized starting documents
- Requests approval — then
/clear and inject FIRST_PROMPT.md as the execution contract
Architecture
plugins/pre-vibe/
├── .claude-plugin/plugin.json # Plugin manifest
├── .mcp.json # stdio MCP server
├── skills/pre-vibe/SKILL.md # Workflow guidance
├── scripts/
│ ├── mcp_server.py # MCP tool surface
│ ├── pre_vibe.py # CLI / compatibility facade
│ ├── pv_*.py # Routing, scan, questions, settings, artifacts
│ └── pyproject.toml # Python project config
└── README.md
The Python layer handles deterministic routing, scanning, validation, and safe writing. Claude Code authors the final task-specific Markdown using MCP tool results and structured workflow data.
Documentation
Safety
- Allowlist project scanning — only reads safe filenames and suffixes
- Skips secrets, keys, tokens, databases, logs
- Keeps output paths inside the active project
- Never silently overwrites existing root
CLAUDE.md
Development
