SkillOpt Skill Writer is a bilingual Codex skill for creating, improving, and validating agent skills with a SkillOpt-inspired workflow: scored evidence, bounded edits, validation gates, rejected-edit memory, and compact deployable SKILL.md artifacts.
SkillOpt 技能写作器 是一个中英文双语 Codex skill,用来创建、改进和验证 agent skill。它借鉴 SkillOpt 论文思想:把 SKILL.md 当作可训练的外部状态,用小步 add/delete/replace 编辑迭代,并通过验证门控决定是否接受修改。
SkillOpt, Codex skill, Claude skill, agent skill, skill writer, skill optimizer, skill creation, SKILL.md, best_skill.md, validation gate, text-space optimizer, rejected-edit buffer, bounded edits, prompt optimization, agent workflow, 写 skill, 创建 Codex 技能, 技能文档, 技能优化, 验证门控, 智能体技能, 工作流转 skill
- Create new
SKILL.mdfiles from repeatable workflows. - Improve existing skills with bounded, auditable edits.
- Validate skill changes with tests, rubrics, or dry-run prompts.
- Keep rejected edits as negative feedback instead of shipping them.
- Produce bilingual documentation and trigger phrases for discoverability.
- 把可复用流程写成新的
SKILL.md。 - 用可审计的小步编辑改进已有 skill。
- 用测试、评分表或 dry run prompt 验证修改。
- 记录被拒绝的编辑,避免重复犯同类错误。
- 输出中英文触发词和说明,让别人更容易搜索到。
Copy this folder into your Codex skills directory:
~/.codex/skills/skillopt-skill-writer/
Windows example:
Copy-Item -Recurse .\skillopt-skill-writer "$env:USERPROFILE\.codex\skills\skillopt-skill-writer"Example prompts:
- "Create a Codex skill for reviewing financial PDFs."
- "Improve this SKILL.md using validation-gated edits."
- "Turn this repeatable workflow into a skill."
- "帮我写一个 Codex skill,用来处理论文精读。"
- "用 SkillOpt 的思路优化这个技能文档。"
skillopt-skill-writer/
SKILL.md
README.md
README.zh-CN.md
LICENSE
topics.txt
references/
skillopt-method.md
validation-gates.md
templates.md
This skill is inspired by the paper "SkillOpt: Executive Strategy for Self-Evolving Agent Skills" and adapts its optimization ideas into a practical writing workflow for Codex-style skills.
It is not the official SkillOpt training framework. It is a lightweight authoring skill for people who want to write better skills with SkillOpt-style discipline.
MIT