其实，

make-goal：把长期 AI 任务从聊天记录里解放出来

摘要

许多 AI agent 在短期任务中表现优异。但当面临长周期项目时质量往往会漂移。从原因在于来看，

• 计划散落在聊天记录中，难以统一管理；
• 环境依赖没有写清楚，验证方式靠临时记忆；怎么说呢，
• 下一轮 agent 接手时又需重新理解上下文。

make-goal 并不是让 AI 写一个更漂亮的计划，而是把复杂目标整理成一个可执行、可恢复、可验证的 goal.md。它将目标、约束、运行环境、里程碑、验收标准和进度记录放进同一个文件，让 Codex、Claude Code 或其他 agent 能直接进入长期执行模式。其实，

使用者痛点一览

• 对话记录太长。 agent 无法一次性读取完整上下文。
• 口头约束容易被遗漏，导致后续执行出现偏差。
• 环境依赖经常只存在于某一轮对话里缺乏持久化描述。
• 验证方式不明确，完成标准随时间变得模糊。
• 进度没有结构化记录，恢复成本高昂。

长任务真正缺的不是计划。而是可执行上下文

说到典型流程，

1. 在聊天里描述一个目标。
2. 让 agent 给出计划。怎么说呢，
3. 让它开始修改代码或整理资料。
4. 中途不断补充上下文。
5. 下一次会话重新解释一遍。

This workflow works fine in minutes but fails over hours or days because “task state” never becomes a stable asset.

聊天记录的问题是：

• 上下文太长，agent 不一定能完整读取；
• 使用者口头约束易被遗漏；
• 环境依赖常驻单轮对话；
• 验证方式不明确，完成标准越来越模糊；
• 进度未结构化记录，恢复成本高。

 的主要判断是：长期任务需要的是一次性 plan，而是一份可以被 agent 反复读取并执行的目标合约。

goal.md 应该包含什么？

A good must answer se questions:

Description of final product or artifact.

---

Name of files/repos/links that must be read first.

---

No-breaking constraints that must be honored.

---

---

---

---

---

---

---

The tool guides agent to ask questions first and generate document only after missing information—especially execution environment—is filled in. It does not rush into a task list;it ensures all prerequisites are documented before proceeding.

Many people think “a clear goal” is enough. In reality it’s often harness— set of runtime dependencies and environment configuration—that holds up an entire project. Without a harness you’re effectively handing your agent a vague instruction set that can misinterpret or fail silently.

A typical harness covers:

• Runtime
• Package manager
• CLI tools
• External services or connectors
• Environment variables and secret names
• Local services or databases and third‑party APIs
• Filesystem & network permissions
• Baseline checks
• Fallback strategy if dependencies are missing

If se aren’t written into your goal file your next agent will have to guess—guessing might work sometimes but more often leads to incorrect assumptions or outright failures. Therefore during goal‑collection phase make‑goal explicitly asks you to fill out a harness section:

md

Execution Harness

Runtime: Package manager: Required CLI tools: Required services: Environment variables: Permissions: Readiness checks: Fallbacks:

This may feel less exciting than a “feature list,” but it determines wher your task can run automatically over weeks or months without human intervention.

The goal file generated by make‑goal isn’t just a read‑me;it defines whole loop an agent follows:

text Read goal.md → Read progress log → Inspect repo state → Pick next smallest useful increment → Implement → Validate → Record progress → Continue

Benefits:

• Reduce drift – every round starts from same source of truth instead of relying on chat context.
• Facilitate recovery – if a run stops mid‑way an incoming agent can simply read goal.md plus progress.log and pick up where it left off.

Make‑goal isn’t tied to one specific project. It’s a generic skill you can drop into any AI environment that supports prompts – Codex skill packages,Claude Code slash commands or any custom prompt engine.

Typical use cases include:

软件重构 & 开源项目初始化 & 产品 MVP & 文档迁移 & 游戏开发 & 知识库清理 & 长期运营任务等场景均适用。说起来，

 

You don't need to clone repos manually;just tell your AI assistant:

text Open https://github.com/talkcozy/make-goal. Read repo,install make-goal in this Codex environment,verify skill is available。n use $make-goal to create a goal.md for this project.

The installation step itself can be delegated back to assistant so you only express intent while everything else gets handled automatically.

A high‑quality goal.md should meet se criteria:

- Self‑contained: understandable without chat history - Executable: next step clear enough - Verifiable: each stage has defined checks - Recoverable: ready if interrupted - Boundary defined: what won’t be done is explicit - Harness detailed: environment and fallback logic included - Progress logging enabled - Completion definition explicit

If you end up with just a task list it's still only a plan—no contract with an AI executor. A complete goal.md reads more like a long‑term agreement than a checklist.

Use make‑goal when your job meets one or more of se signals:

• Task will last days/weeks/months • Spans multiple directories or repositories • Requires repeated verification steps • May switch agents or sessions midway • Environment dependencies are complex • You want maximum autonomous progress from an assistant • Completion criteria extend beyond simple screenshots/one function

In short—if writing code directly feels rushed and uncertain before starting work n make‑goal gives you peace of mind by clarifying everything upfront.

AI agents grow stronger yet y also demand clearer boundaries. A well-crafted goal.md is not restrictive;rar it builds an enduring foundation upon which agents can reliably operate over time. The harder you invest upfront in structuring goals and harnesses,smoor—and seemingly effortless— downstream work will appear.

Core takeaway from make‑goal:

---