vibe-agents
GitHub用于生成AGENTS.md及AI配置文件的技能,指导AI助手构建MVP。通过加载PRD和技术设计上下文,询问用户工具偏好与技术水平,实例化模板并填充占位符,确保AI编码流程的连续性与标准化。
Trigger Scenarios
Install
npx skills add KhazP/vibe-coding-prompt-template --skill vibe-agents -g -y
SKILL.md
Frontmatter
{
"name": "vibe-agents",
"description": "Generate AGENTS.md and AI configuration files for your project. Use when the user wants to create agent instructions, set up AI configs, or says \"create AGENTS.md\", \"configure my AI assistant\", or \"generate agent files\".",
"allowed-tools": "Read, Write, Glob, Grep, AskUserQuestion"
}
Vibe-Coding Agent Configuration Generator
You are helping the user create AGENTS.md and tool-specific configuration files. This is Step 4 of the vibe-coding workflow.
Your Role
Generate the instruction files that guide AI coding assistants to build the MVP. Use progressive disclosure - master plan in AGENTS.md, details in agent_docs/.
Session Continuity
- Keep Step 4 outputs aligned with prior PRD and Tech Design context.
- If prior chat context is missing, require a compact handoff summary before generating files.
- Add continuity hints in generated instructions so users avoid empty-chat resets during Step 5.
Naming Policy
Use model family names in examples and recommendations unless the user explicitly requests pinned versions.
Prerequisites
- Look for
docs/PRD-*.md- REQUIRED - Look for
docs/TechDesign-*.md- REQUIRED - If either is missing, suggest running the appropriate skill first
Step 1: Load Context
Handoff Context block? Read it first. If the Tech Design (or PRD) ends with a ## Handoff Context block, pre-fill the app name, user level, chosen stack, and AI coding tool from it, confirm them in one line ("Got it from the handoff: building X with Y, using Z"), and don't re-ask what it already answers in Step 2. No Handoff Context block? Just ask — older documents won't have it.
Extract from documents:
From PRD:
- Product name and description
- Primary user story
- All must-have features
- Nice-to-have and excluded features
- Success metrics
- UI/UX requirements
- Timeline and constraints
From Tech Design:
- Complete tech stack
- Project structure
- Database schema
- Implementation approach
- Deployment platform
- AI tool recommendations
Step 2: Ask Configuration Questions
Ask the user:
Which AI tools will you use? (Select all that apply)
- Claude Code (terminal-based)
- Gemini CLI (free terminal agent)
- Google Antigravity / equivalent (agent-first IDE)
- Cursor (AI-powered IDE)
- VS Code + GitHub Copilot
- Lovable / v0 (no-code)
Then ask:
What's your technical level?
- A) Vibe-coder
- B) Developer
- C) In-between
Step 3: Instantiate the Templates
This repository ships the canonical templates in /templates/. Read and instantiate them — do NOT write your own versions of these files.
Copy ALL of these into the project root:
| Template | Destination |
|---|---|
templates/AGENTS.md |
AGENTS.md |
templates/MEMORY.md |
MEMORY.md |
templates/REVIEW-CHECKLIST.md |
REVIEW-CHECKLIST.md |
templates/agent_docs/*.md (all five files) |
agent_docs/ |
Then fill every placeholder using the PRD and Tech Design. Templates use exactly two placeholder kinds:
[REPLACE: description]— fill in with project-specific content.[CHOOSE: option A | option B | option C]— pick the ONE matching option and delete the rest.
No square-bracket placeholders may remain when you're done.
Key fills:
- AGENTS.md — Overview & Stack, Setup & Commands from the Tech Design. Tune the existing behavioral sections (
How I Should Think,What NOT To Do,Engineering Constraints) to the user's level — fill, don't create. Fill Roadmap Phase 2 with the PRD's must-have features. Set Current State to "Project setup — nothing built yet". - MEMORY.md — Initialize
## 🏗️ Active Phase & Goalfrom the PRD's Phase 1. - REVIEW-CHECKLIST.md — Copy as-is; it has no placeholders.
- agent_docs/tech_stack.md — Every library, version, and setup command from the Tech Design, plus short canonical code examples.
- agent_docs/code_patterns.md — Resolve the CHOOSE lists (architecture pattern, data fetching, file naming) from the Tech Design.
- agent_docs/project_brief.md — Product vision and conventions.
- agent_docs/product_requirements.md — Complete feature list (MoSCoW), user stories, and success metrics from the PRD.
- agent_docs/testing.md — Test frameworks and commands per the Tech Design.
Resulting structure:
project/
├── AGENTS.md # Master plan
├── MEMORY.md # Session memory
├── REVIEW-CHECKLIST.md # Definition of done
├── agent_docs/
│ ├── tech_stack.md # Tech details
│ ├── code_patterns.md # Code style
│ ├── project_brief.md # Persistent rules
│ ├── product_requirements.md # PRD summary
│ └── testing.md # Test strategy
├── CLAUDE.md # If Claude Code selected
├── .cursor/rules/vibe.mdc # If Cursor selected
├── .agent/rules/vibe.md # If Antigravity selected
└── (Codex needs nothing — it reads AGENTS.md natively)
Step 4: Generate Tool Configs
AGENTS.md is the universal contract — Codex reads it natively, and most modern agents do too. Tool configs are thin adapters that point at it. Ready-to-copy versions live in templates/tool-adapters/.
| Tool | File | What it is |
|---|---|---|
| Claude Code | CLAUDE.md |
3-line pointer to AGENTS.md + agent_docs/ |
| Codex | — none — | Nothing; AGENTS.md is Codex's native instruction file. Optional: ~/.codex/prompts/ for personal slash prompts |
| Antigravity | .agent/rules/vibe.md |
Always-on workspace rule pointing at AGENTS.md |
| Cursor | .cursor/rules/vibe.mdc |
Rule with alwaysApply: true pointing at AGENTS.md |
| Any other tool | its custom-instructions feature | "Read AGENTS.md — it is the source of truth for this project." (Gemini CLI users: point it at AGENTS.md the same way.) |
CLAUDE.md (Claude Code)
# CLAUDE.md
Read AGENTS.md first. It is the source of truth for this project: roadmap, commands, rules.
Implementation details live in `agent_docs/` — consult them before coding.
Plan before coding, build one feature at a time, verify before moving on.
.cursor/rules/vibe.mdc (Cursor)
Legacy .cursorrules still loads but is deprecated — delete it if one exists.
---
description: Vibe-coding project rules — source of truth is AGENTS.md
alwaysApply: true
---
Read AGENTS.md first. It is the source of truth: roadmap, commands, rules. Details live in `agent_docs/`.
- Plan before coding; get approval, then build one feature at a time.
- Don't act as a manual linter — rely on the project's configured formatter/linter; don't reformat files you didn't touch.
- Never delete files or change the database schema without confirmation.
.agent/rules/vibe.md (Antigravity)
Current Antigravity reads AGENTS.md natively; this always-on workspace rule reinforces it. Global rules live in ~/.gemini/GEMINI.md.
Read AGENTS.md first. It is the source of truth for this project: roadmap, commands, rules.
Implementation details live in `agent_docs/` — consult them before coding.
Plan before coding, build one feature at a time, verify before moving on.
After Completion
Write all files to the project, then tell the user:
Files Created:
AGENTS.md- Master planMEMORY.md- Session memoryREVIEW-CHECKLIST.md- Definition of doneagent_docs/- Detailed documentation- [Tool-specific configs based on selection]
Project Structure:
your-app/ ├── docs/ │ ├── research-[App].md │ ├── PRD-[App]-MVP.md │ └── TechDesign-[App]-MVP.md ├── AGENTS.md ├── MEMORY.md ├── REVIEW-CHECKLIST.md ├── agent_docs/ │ ├── tech_stack.md │ ├── code_patterns.md │ ├── project_brief.md │ ├── product_requirements.md │ └── testing.md └── [tool configs]Next Step: Run
/vibe-buildto start building your MVP, or say "Build my MVP following AGENTS.md"
Version History
-
06d0796
Current 2026-07-19 18:25
引入Handoff Context块实现阶段间无缝衔接,消除重复提问;重写部分4为单一模板实例化工作流,规范[REPLACE:]/[CHOOSE:]占位符;移除冗余变体,强化Claude Code适配及合同检查机制。
- e063721 2026-07-05 14:58


