About PrePrompt

PrePrompt is a Pre-AI Cognitive Layer—a structured thinking protocol you pass through before sending a request to an AI system. It is designed to reduce token usage, prevent AI over-generation, and give you back control over your AI outputs by forcing you to structure your intent into five distinct stages.

How to use PrePrompt

1

Intent Lock

Define the desired end-state

Guidelines

Define the desired end-state clearly and specifically.

Tips for a better prompt

  • State WHAT should exist when the AI is done — not just what it should do.
  • Each bullet should be independently verifiable (can you test it?). If not, it's still too vague.
  • Avoid words like 'good', 'clean', 'proper'. Use measurable terms instead.
2

Reality Anchor

Describe current system state

Guidelines

Describe your current system state explicitly.

Tips for a better prompt

  • List exact library versions (e.g. 'Next.js 14.2.3'). Version differences change behavior.
  • Name the exact files and directories the AI should read, modify, or leave untouched.
  • State what DOESN'T exist yet — the AI often hallucinates pre-existing code.
3

Constraint Cage

Define non-negotiable boundaries

Guidelines

List every non-negotiable boundary.

Tips for a better prompt

  • Use 'MUST NOT' and 'NEVER' for hard prohibitions — these carry far more weight than 'avoid' or 'prefer not'.
  • Think adversarially: what would a technically correct but wrong solution look like? Forbid that.
  • Separate style constraints (formatting, naming) from architectural constraints (no new deps, specific patterns).
4

Action Slice

Smallest meaningful execution unit

Guidelines

Define the ONE thing the AI should do right now.

Tips for a better prompt

  • One action slice = one commit. If it would be two commits, it's two slices.
  • End with 'Nothing else.' — this is surprisingly effective at stopping scope creep.
  • Think about dependencies: only ask for things the current codebase can already support.
5

Response Contract

Specify output format requirements

Guidelines

Specify exactly how the AI should format its response.

Tips for a better prompt

  • Specify the exact format: unified diff, JSON object, markdown table, numbered list, etc.
  • State what to OMIT: 'no preamble', 'no explanations', 'no closing remarks' saves you from skimming.
  • If you want reasoning, ask for it in <thinking> tags — this keeps the final output clean.

Developer Info

Email: inkyung.huh.ink@gmail.comGitHub Repository: https://github.com/inxnxng/PrePromptBlog: https://jadelog.vercel.app/