Final State Delivery — AI skill for Claude Code

Final-State Delivery

[English](README.md) | [简体中文](README.zh-CN.md)

[](https://github.com/Leroyha/Final-state-delivery/stargazers) [](#compatibility) [](https://opensource.org/licenses/MIT)

AI agents should deliver the result, not the history of the conversation

Final-State Delivery is a concise specification and practical documentation set for AI coding agents used in Vibe Coding workflows. It addresses a recurring failure mode: a rejected idea, abandoned implementation, or local correction leaks into the final codebase as commentary, tests, UI text, memory, or a new global rule.

**The Tomato Egg Without Dongpo Pork Principle**

When the intended result is a simple dish, feature, page, or document, deliver that result. Do not turn the discarded ingredient or discarded implementation into a permanent part of the product's identity.

This repository gives teams a compact `AGENTS.md` rule, a scope model, a set of tool-neutral practices, case studies, and an evaluation rubric. It is designed for ChatGPT/Codex, Claude Code, Gemini CLI, GitHub Copilot, Grok Build, Cursor, and other instruction-following coding agents.

The problem in one minute

In an ordinary engineering workflow, the product records the supported behavior and the rationale for durable constraints. In an AI-assisted workflow, the model sees the entire working conversation and may treat every salient statement as a candidate instruction. A local correction can therefore reappear as:

• a PR title that describes what was removed instead of what now exists;
• comments, tests, or validators for an abandoned approach;
• UI or documentation text that advertises an absent feature;
• a memory entry that tu