ai-native-builder-workflow

name	ai-native-builder-workflow
description	A complete end-to-end framework for non-technical product managers to build and ship software using AI coding agents. Use this when starting a side project, building a prototype, or automating internal tools without an engineering team.

This workflow enables non-technical individuals to build production-ready applications by orchestrating AI models as a technical co-founder, developer, and QA lead.

Core Principles

• The CTO Persona: Treat the AI as a Technical Owner. Instruct it to challenge your ideas, avoid "people-pleasing" (sycophancy), and own the technical architecture while you own the problem and user experience.
• Exposure Therapy: Gradually move from simple chat interfaces (ChatGPT/Claude Projects) to dedicated builders (Bolt/Lovable) to pro-level IDEs (Cursor/Claude Code).
• Planning over Vibe-ing: Never let the AI start coding until a markdown plan is finalized. Eager coding leads to architectural debt and complex bugs.

The /Command Workflow

Implement these custom prompts as reusable /commands within your AI coding environment (Cursor, Claude Code, or IDE system prompts).

1. Capture: /create-issue

Purpose: Quickly capture bugs or features without breaking development flow.

• Instruction: Tell the AI to stop what it's doing and summarize the thought into a specific format.
• Format: TLDR, Current State, Expected Outcome, and Priority.
• Integration: Use MCP (Model Context Protocol) to automatically create a ticket in Linear or GitHub.

2. Deep Dive: /exploration-phase

Purpose: Force the AI to understand the technical implications before writing code.

• Process: Provide the issue ID as context.
• Requirement: The AI must analyze the codebase and ask 5-10 clarifying questions regarding data models, UX, edge cases, and architectural impact.
• Goal: Identify "Key Areas" of the code that will be affected.

3. Strategy: /create-plan

Purpose: Generate a source-of-truth roadmap for the build.

• Structure: Create a plan.md file including:
  • TLDR: High-level goal.
  • Critical Decisions: Tech stack choices or logic changes.
  • Task Checklist: Step-by-step implementation guide with status checkboxes ([ ]).
• Review: Manually approve this plan before moving to execution.

4. Execute: /execute

Purpose: Move the plan into code.

• Process: Feed the plan.md to the coding agent (e.g., Cursor Composer or Claude Code).
• Control: Execute one task at a time to ensure the UI and logic remain stable.

5. Multi-Model QA: /peer-review

Purpose: Catch errors that a single model might miss by creating "model friction."

• Technique: Have different LLMs review the same code.
• Workflow:
  1. Run /review with Claude to find its own mistakes.
  2. Copy the code into a different model (e.g., GPT-4o or Gemini 1.5 Pro).
  3. Use the /peer-review prompt: "You are the dev lead. Other team leads found these issues [paste issues]. Either fix them or explain why they are not real issues based on our specific context."
• The "Fight": Allow the models to argue technical points until a consensus is reached.

6. Continuous Learning: /learning-opportunity

Purpose: Build your technical intuition while building.

• Prompt: "I am a technical PM in the making. Explain this specific technical decision or error using the 80/20 rule. Focus on architecture and mental models, not just syntax."

Maintaining the "Harness"

To keep the AI effective as the project grows, you must maintain its documentation.

• /update-docs: After every major feature, have the AI update the project’s documentation (e.g., architecture.md, api-routes.md) so the next agent session has full context.
• Post-Mortems: When the AI makes a mistake, ask: "What in your system prompt or current documentation caused this error?" Update the system prompt to prevent that specific category of error from recurring.

Examples

Example 1: Feature Ideation

• Context: You want to add a drag-and-drop "fill-in-the-blank" quiz type to a learning app.
• Input: /create-issue I want a drag and drop quiz type. 30% of tests should have this. 6 potential answers, 2 blanks.
• Application: AI creates a Linear ticket. You then run /exploration-phase where the AI asks how the state should be handled if a user drags the same answer to two different spots.
• Output: A comprehensive technical plan that accounts for drag-and-drop library dependencies before any code is written.

Example 2: Bug Resolution

• Context: The app crashes only on mobile Safari.
• Input: Run the code through GPT-4o for a second opinion.
• Application: GPT identifies a CSS incompatibility. Use /peer-review to feed that feedback back to Claude.
• Output: Claude acknowledges the oversight and provides a cross-browser compatible fix.

Common Pitfalls

• The Sycophancy Trap: The AI will often agree with your bad ideas just to be helpful. Explicitly prompt it to be a "Cantankerous CTO" who protects the codebase.
• The "Slop" Accumulation: Letting the AI generate thousands of lines without review. Use a "deslop" mindset: ask the AI to refactor for conciseness and remove redundant comments or unused imports after a feature is done.
• Skipping the /review: Assuming the code works because it "looks" right. Always run the code locally and trigger a /review from a competing model.