让 AI 助手从真实任务里持续变好，而不是靠空泛的“自我进化”口号。
适合谁 / For: 希望 Codex 或兼容 AI 助手少犯重复错误的个人开发者、团队和 AI 工作流维护者。 / Individual builders, teams, and AI workflow maintainers who want Codex or a compatible AI assistant to stop repeating avoidable mistakes.
得到什么 / Outcome: 把任务证据、用户纠正和重复模式整理成一份可审阅的经验、规则或 Skill 升级提案，为下一次工作提供更可靠的默认动作；不会让 AI 自主改写目标。 / A reviewable lesson, rule, or skill-upgrade proposal grounded in task evidence, user corrections, and recurring patterns—giving future work safer defaults without letting the AI rewrite its own goals.
安装后，在一个有实际证据的任务结束时直接说：
After installation, use this prompt at the end of a task with real evidence:
Use $ai-evolution-flywheel to capture this task's reusable lesson, choose micro, review, or upgrade, and propose the next default without writing to long-term memory unless I approve.
你会得到一份简短、可审阅的复盘或升级提案；它不会替你自动修改长期规则。 / You receive a compact, reviewable lesson or upgrade proposal; the skill does not automatically change durable rules for you.
源码状态 / Source status: This repository currently provides rolling source from
main, not a fixed or stable release. Until a versioned Git tag and matching GitHub Release exist, pin a reviewed commit if reproducibility matters and inspect changes before updating. / 本仓库目前提供的是main分支滚动源码，不是固定版本或稳定安装版。在出现带版本号的 Git Tag 及对应 GitHub Release 之前，如需可复现使用，请固定到已审阅的 commit，并在更新前检查变更。
隐私与授权 / Privacy and authorization: Store only abstract, reusable lessons—not full chats, credentials, customer data, payment details, or raw production logs. The skill never grants permission to deploy, push, publish, pay, change permissions, access secrets, delete data, or make public commitments; those actions still require explicit human approval. / 只保存抽象、可复用的经验，不保存完整聊天、凭据、客户数据、支付信息或原始生产日志。本 Skill 不会授予部署、推送、发布、付费、改权限、读取密钥、删除数据或作公开承诺的权限；这些操作仍须人工明确授权。
Most AI assistants do not really learn from the work you do together. They may remember fragments, but the useful lessons often disappear:
- the user corrected a recurring mistake;
- a project had a safer startup path;
- a deployment or deletion boundary became clearer;
- a repeated workflow should become a default;
- a failed attempt revealed a better verification step.
This skill gives those moments a simple operating system.
It is not designed for uncontrolled recursive self-improvement. It is designed for builders, teams, and solo operators who want an AI assistant that gets more reliable over time while keeping human control, privacy, and verification intact.
For the deeper theory, research basis, and product roadmap, read:
The short version: useful AI evolution should be a controlled work system, not a vague promise that an agent will magically improve itself. It needs task traces, layered memory, reflection, external evaluation, rule compilation, skill upgrades, and clear human approval boundaries.
micro: one task -> one lesson -> next default
review: many tasks -> repeated patterns -> rule/skill candidates
upgrade: stable pattern -> AGENTS/SOP/template/script/skill proposal
The skill asks the agent to choose the smallest mode that fits the evidence.
For a single task, it closes with:
What evidence do we have?
What worked?
What was wasteful?
What should be the default next time?
Should this be stored, and where?
ai-evolution-flywheel helps Codex or another compatible AI agent:
- capture reusable lessons after a task;
- classify lessons as facts, judgment patterns, project methods, risk boundaries, expression preferences, tool rules, or open questions;
- avoid saving full chat transcripts and sensitive raw material;
- produce structured memory candidates when direct knowledge-base writes are unavailable;
- keep project logs separate from cross-project reusable knowledge;
- require external verification instead of trusting AI self-review alone;
- preserve human approval for high-risk actions such as deployment, public publishing, payment, permissions, secrets, and deletion.
Use this skill when a task produces:
- a user correction that should affect future behavior;
- a repeated mistake or repeated waste;
- a project workflow that should become a template;
- a safety, privacy, deployment, deletion, or permission boundary;
- a writing, communication, or output-format preference;
- a tool-routing rule or validation habit;
- a knowledge-base candidate that should be reviewed before storage.
You can also make it a default closeout skill for every substantial Codex task.
.
├── README.md
├── LICENSE
├── CONTRIBUTING.md
├── docs/
│ ├── AI进化方向重设计_2026-07-05_v1.md
│ └── AI进化论参考文档_2026-07-05_v1.md
└── skills/
└── ai-evolution-flywheel/
├── SKILL.md
├── agents/
│ └── openai.yaml
├── references/
│ ├── global-settings.md
│ ├── install-and-share.md
│ └── methodology.md
└── scripts/
└── record_evolution.py
Clone or download the repository, review the source commit, then copy the skill folder into your local Codex skills directory:
cp -R skills/ai-evolution-flywheel ~/.codex/skills/Restart Codex so it can discover the new skill.
This installs a local copy of the repository's rolling source. It is not a versioned package install. Record or pin the commit you reviewed when repeatability matters.
Invoke it directly:
Use $ai-evolution-flywheel to capture this task's reusable lessons and decide whether this is micro, review, or upgrade.
Or add a global rule to your Codex instructions:
At the end of each meaningful task, if reusable lessons, user corrections, workflow changes, risk boundaries, or tool rules appear, use `ai-evolution-flywheel` to capture:
- evidence;
- what worked;
- what was wasteful;
- what should be the default next time;
- whether it should be stored.本轮进化记录：
- 证据：README、执行日志和测试输出已检查。
- 有效做法：先读取项目入口文件，再决定是否修改代码。
- 浪费点：重复搜索了已经写在 execution-log 里的信息。
- 下次默认动作：接手项目时先读 README、AGENTS 和 docs/execution-log.md。
- 是否入库：可入库，建议写入项目方法。When direct memory storage is not available, the skill can output a reviewable candidate:
可入库候选
- 来源项目：
- 线程标题：
- 内容类型：事实 / 判断方式 / 项目方法 / 风险边界 / 表达偏好 / 工具规则 / 待确认问题
- 提炼结果：
- 建议写入位置：
- 是否敏感：
- 是否待确认：
- 不应保存的原文：The bundled script creates a compact Markdown evolution record:
python3 skills/ai-evolution-flywheel/scripts/record_evolution.py \
--mode micro \
--source-project "my-project" \
--thread-title "Fix onboarding flow" \
--content-type "项目方法" \
--evidence "README, AGENTS, and execution log reviewed" \
--effective "Read project docs before editing" \
--waste "Repeated an already-documented check" \
--next-action "Start future project recovery from README, AGENTS, and execution log" \
--storage "可入库候选" \
--output ./evolution-record.mdReview the generated file before copying it into any long-term memory system.
The AI can suggest lessons, but humans keep control over direction, permissions, public commitments, and high-risk operations.
Memory should be grounded in task results, tests, user feedback, source citations, screenshots, logs, or other external checks. AI self-review is not enough.
Store abstract reusable lessons, not raw private material. Do not store credentials, API keys, cookies, customer data, contract details, payment information, or raw production logs.
Good evolution records are short. One useful default action is better than a long archive nobody reads.
This is not:
- a model fine-tuning system;
- a vector database;
- a private-data collection tool;
- an autonomous self-modifying agent;
- a replacement for tests, review, or human judgment.
It is a lightweight workflow skill that makes AI-assisted work accumulate practical memory safely.
MIT License. Free for personal, commercial, research, and educational use.
Created and maintained by 森海大林.
Contributions, forks, translations, and adaptations are welcome.