Agent Stack Picker

Recommend one stack that closes as much of the product as possible with battle-tested solutions.

This skill is designed for environments where:

• AI agents still overestimate their architecture and implementation quality.
• Freshly generated code is often untrusted until humans and tests validate it.
• Low-experience operators may not be able to correct a weak technical plan.

How It Works

1. Decide whether the problem should be solved by building an app at all.
2. Ask a short quiz only if the prompt does not already contain enough constraints.
3. Separate the product into core business logic, UX, commodity capabilities, AI capabilities, and operational requirements.
4. Choose the highest-leverage solution for each layer with a strong bias toward battle-tested products and platforms.
5. Return one recommended stack, explain every part, then mention brief alternatives.

Files To Read

Read these files progressively:

• references/principles.md
• references/defaults.md
• references/quiz.md
• references/non-app-paths.md

Do not load everything if only one file is needed.

What To Optimize For

• minimal agent-written code;
• minimal integration glue;
• battle-tested products and frameworks;
• TypeScript-friendly stacks;
• realistic self-hosting complexity;
• one clear recommendation instead of a long package dump.

Output Requirements

Always:

• recommend one primary stack;
• explain each component as product, platform, framework, service, or custom code;
• explain what stays custom and why;
• say why the stack is safer for agent-driven development;
• include brief alternatives with explicit tradeoffs.

Short Quiz

Ask only the missing items:

1. What must stay internal or self-hosted, and what may use external services?
2. What kind of product is this?
3. Which parts must be deeply custom because they define the business or UX?
4. Are there compliance, region, procurement, or data restrictions?
5. Are there existing stack constraints, especially around TypeScript, frontend, cloud, or database?

Present Results To User

Use this structure:

1. Problem framing
2. Missing assumptions
3. Recommended stack
4. Layer-by-layer explanation
5. What stays custom and why
6. Why this is safer for agent-driven development
7. Alternatives

Troubleshooting

• If the request is too vague, use only the short quiz and avoid a long discovery session.
• If the user asks for one library, widen the question to the full stack and system layers first.
• If no app should be built, say that clearly instead of forcing a stack recommendation.