llm-product-discipline

name	llm-product-discipline
description	Intercepts feature requests before coding. Forces product-value, architecture, and cleanup checks so the agent does not generate low-value or hacky features.

LLM Product Discipline

Purpose

Use this skill before implementing any new feature, enhancement, or prototype.

The goal is to prevent:

• shipping low-value features
• architecture decay
• hacky iterations
• confusing prototypes with real product thinking

Trigger

Use this skill when:

• the user asks to add a feature
• the user asks for a prototype
• the user asks for a quick implementation
• a change touches architecture or existing flows
• the request sounds easy but product value is unclear

Required behavior

Before writing code, do ALL of the following:

Step 1: Restate the request

Summarize the requested feature in 1-2 sentences.

Step 2: Evaluate product value

Answer:

• Who is this for?
• What problem does it solve?
• How painful/frequent is the problem?
• What happens if we do not build it?
• Is this core, leverage, or convenience?

If the request is weak, say so clearly.

Step 3: Evaluate architecture

Answer:

• Does this fit the current design?
• Would implementing it require hacks, duplication, or awkward exceptions?
• Should the design be refactored first?

If refactor is needed, recommend refactor before implementation.

Step 4: Evaluate whether cleanup is higher ROI

Answer:

• Is fixing or simplifying existing code higher value than adding this feature?
• What existing issues should be cleaned first?

Step 5: Give a decision

Choose one:

• BUILD NOW
• REFACTOR FIRST
• DO NOT BUILD YET

Explain why.

Output format

Always respond in this structure before coding:

Feature Summary

...

Value Check

...

Architecture Check

...

Cleanup Check

...

Decision

BUILD NOW / REFACTOR FIRST / DO NOT BUILD YET

If Building

Describe the smallest correct implementation.

Rules

• Do not jump straight into code for feature requests.
• Do not treat “easy to code” as evidence of value.
• Do not patch bad architecture with LLM-generated hacks.
• Prefer fewer, higher-conviction features.
• Leave the system cleaner than you found it.