vibe-coding-prompt-template

flowchart LR
    A[💡 Idea] --> B[📊 Research] --> C[📋 PRD] --> D[🏗️ Tech Design] --> E[🤖 AGENTS.md] --> F[🚀 MVP / AI App]

    style A fill:#667eea,stroke:#667eea,color:#fff
    style B fill:#764ba2,stroke:#764ba2,color:#fff
    style C fill:#f093fb,stroke:#f093fb,color:#fff
    style D fill:#4facfe,stroke:#4facfe,color:#fff
    style E fill:#00f2fe,stroke:#00f2fe,color:#000
    style F fill:#43e97b,stroke:#43e97b,color:#000

This step gives you a quick read on demand, competitors, and whether the scope looks realistic.

1. Open part1-deepresearch.md and copy all of its contents.
2. Paste it into your preferred AI platform Chat (like Claude.ai, ChatGPT, or Gemini) and press Enter.
3. The AI will ask you a few questions about your idea. Answer them truthfully in the chat.
4. The AI will generate a comprehensive research document based on your answers.
5. Save the output into a local file named research-[YourAppName].md (or .txt) or simply keep this chat open for Step 2.

Tip: if your chat tool supports web search, source grounding, URL context, or deep research mode, turn it on and require cited claims with access dates.

This turns the rough idea into a scope you can build against.

1. Copy the contents of part2-prd-mvp.md.
2. Option A (Same Chat): If you kept your chat open, paste the prompt right below the Deep Research output.
3. Option B (New Chat): Start a fresh chat, paste your saved research-[YourAppName].md content, and then paste the Part 2 prompt below it.
4. Press Enter, answer any clarifying questions the AI asks, and let it generate your requirements.
5. Save the final output as PRD-[YourAppName]-MVP.md.

This step helps you choose the stack, deployment target, AI provider strategy if the product needs AI, and the verification path.

1. Copy the contents of part3-tech-design-mvp.md.
2. Paste it into your ongoing conversation (or into a new one, making sure to attach the PRD-[YourAppName]-MVP.md from Step 2 as context).
3. The AI will ask questions regarding your budget, timeline, and complexity tolerance.
4. Discuss the trade-offs it presents, including no-code/full-code, Vercel vs. Cloudflare, and whether AI evals are required.
5. Once a stack is decided, save the output as TechDesign-[YourAppName]-MVP.md.

CLI shortcut: tell your AI agent to run npx vibeworkflow from your project folder — the CLI is agent-driven, not meant to be run by hand. If you already have docs/PRD-*.md + docs/TechDesign-*.md (with the JSON meta block), it scaffolds the files below (skipping anything you've already edited) and verifies with npx vibeworkflow doctor. If the docs are missing, it installs the planning skills, auto-detects your AI tools, and prints agent instructions that drive the research → PRD → Tech Design interviews for you — your agent asks the questions one at a time; you just answer. The paste flow below still works everywhere.

This step fills out AGENTS.md and the supporting docs from your PRD and tech design.

1. Click "Use this template" in GitHub (or clone this repository locally).
2. Open this cloned repository folder in your AI IDE (like Cursor or VS Code).
3. Create a docs/ folder in your project root if it does not already exist.
4. Move your saved documents into docs/ using these names:
   • docs/PRD-[YourAppName]-MVP.md
   • docs/TechDesign-[YourAppName]-MVP.md
   • optional: docs/research-[YourAppName].md (or .txt for backward compatibility)
5. Open the AI Chat inside your IDE, type: "Read part4-notes-for-agent.md, follow its instructions, and set up my workspace."
6. The agent should copy the boilerplates from /templates/, generate selected tool configs (CLAUDE.md, .cursor/rules/, GEMINI.md, .codex/config.toml, .agents/skills/, etc.), and fill placeholders using the files in docs/.

Default generated files:

• AGENTS.md
• MEMORY.md
• REVIEW-CHECKLIST.md
• agent_docs/project_brief.md
• agent_docs/tech_stack.md
• agent_docs/testing.md

Optional generated files:

• agent_docs/code_patterns.md when the codebase has real conventions to preserve.
• agent_docs/product_requirements.md when the PRD is long enough to need a build-facing summary.
• .claude/, .cursor/, .github/, .codex/, .gemini/, and local-agent files only for tools the user selected.
• agent-permissions.example.json only when AI tools, MCP servers, or product actions are in scope.

Choose your development environment and start iterating:

1. Ensure your newly generated AGENTS.md, agent_docs/, and tool configuration files are physically in the project folder.
2. Give your agent its first command:

   "Read AGENTS.md, propose a Phase 1 plan, wait for my approval, and then build it step by step."

3. Treat the agent like a junior developer. Ask it to stop after each major feature, explain the diff, run the project-specific verification commands, and use browser/mobile checks for user-visible flows.
4. Repeat the loop until your MVP is complete:

Recommended Loop:

╭──────────────╮      ╭──────────────╮      ╭──────────────╮
│   📝 Plan    │ ───>│  ⚡ Execute │ ───>│  🔍 Verify  │
│  (Approve)   │      │  (One Feat)  │      │    (Test)    │
╰──────────────╯      ╰──────────────╯      ╰──────────────╯
       ▲                                           │
       └───────────────────────────────────────────┘

To avoid context overload, let the agent write durable project facts into files instead of trying to keep everything in one giant chat:

• Compaction and handoffs: Use native compaction/summarization where the tool supports it. When you switch sessions, have the agent write a specs/001-feature.md or recap.md and load only that file into the new chat.
• Repo-owned memory: Keep decisions and current state in MEMORY.md; tool-side memories are personal and should not replace versionable project docs.
• Cursor/Codex/Gemini context: Use project rules, skills, or GEMINI.md as concise pointers to AGENTS.md and agent_docs/, not as huge prompt dumps.
• If you must restart, attach AGENTS.md, docs/PRD-[YourAppName]-MVP.md, docs/TechDesign-[YourAppName]-MVP.md, and your latest handoff artifact.

• Subagents first: Use focused subagents for research, code review, debugging, and test verification. Use experimental team-style coordination only when agents truly need to communicate or split disjoint modules.
• Plan before edit: Use the tool's actual plan/approval mode where available, then require a short plan before multi-file changes.
• Scoped rules and skills: Keep AGENTS.md as the cross-tool source of truth, then add .cursor/rules/, .claude/agents/, .codex/config.toml, .agents/skills/, or GEMINI.md only as concise tool-specific adapters.
• Task routing: Use Agent tooling compatibility to decide when to use Codex, Claude Code, Cursor, Copilot, Antigravity, local agents, or builder tools.

Use model families instead of pinned version names. It ages better as models get swapped underneath you.

When an agent ignores instructions or behaves inconsistently:

1. Check which instructions/rules/hooks were loaded.
2. Confirm tool permissions and blocked actions.
3. Verify the active session context was not reset.
4. Re-run with explicit instruction order: "Read AGENTS.md, then agent_docs/, then execute."