skill-authoring
Create or update portable Agent Skills under .agents/skills with concise SKILL.md instructions and optional scripts, references, or assets. Use when a user wants skill-creator behavior that works across Codex, Gemini, Claude Code, OpenCode, or other agents without relying on one vendor's built-in creator.
- License 许可
- MIT
- 最近更新
- 2026-11-03
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- skill-authoring
- description
- Create or update portable Agent Skills under .agents/skills with concise SKILL.md instructions and optional scripts, references, or assets. Use when a user wants skill-creator behavior that works across Codex, Gemini, Claude Code, OpenCode, or other agents without relying on one vendor's built-in creator.
03 SKILL.md 正文
<!-- agent-workbench: managed portable-skill -->
Skill Authoring
Dispatch
Use the tier, delivery boundary, and escalation rule for skill-authoring in AI_AGENT_GUIDE.md's Skill Model and Reasoning Routing table, and resolve the tier through the active platform binding.
Workflow
- Read
.agents/prompts/create-agent-skill.mdif present. - If the active environment provides a compatible built-in skill creator, prefer it while keeping the output in the standard project-local format below; otherwise run this portable workflow directly.
- Gather concrete trigger examples and the repeated task the skill should improve.
- Create
.agents/skills/<skill-name>/SKILL.md. - Use frontmatter with only
nameanddescriptionunless a target explicitly requires more. - Add
scripts/,references/, orassets/only when they materially reduce repeated work or increase reliability. - Test scripts and validate frontmatter.
Quality Bar
Keep skills concise, non-obvious, reusable, project-safe, and actionable.
04 同仓库更多 Skills
sync-agent-workbench
Synchronize the fixture workbench.
sync-agent-workbench
Synchronize, audit, or repair composable agent-workbench policy in a project. Use when a user wants canonical AI_AGENT_GUIDE.md, thin vendor loader entrypoints, project-selected profiles, portable prompts, and standard Agent Skills without depending on a marketplace, plugin, global config, or submodule.
math-pdf-reader
Read, transcribe, summarize, or verify mathematical PDFs with high fidelity by combining rendered-page inspection, cautious text extraction, version-matched source checks, and explicit uncertainty. Use for theorem, proof, formula, notation, attribution, or bibliographic work where ordinary PDF text extraction can silently corrupt mathematical content.
loop-until-done
Keep working on a requested task through bounded work, root-cause diagnosis, verification, and retry iterations until explicit completion criteria are met or a maximum iteration limit is reached. Use for Ralph-style self-correction without requiring Claude plugins, Codex plugins, Gemini extensions, or hook support.
linus-review
Run a strict Linus-style code review focused on correctness, backward compatibility, simplicity, data structures, special cases, hardcoded values, root-cause proof, maintainability, and verification. Use when the user asks for Linus mode, Linus review, a blunt maintainer review, harsh review, kernel-style review, or an especially skeptical code review.
integrate-chatgpt-conversation
Retrieve and integrate the current accessible branch of a user-supplied ChatGPT conversation mention or direct ChatGPT URL. Use when a task depends on the full live transcript, including when the user asks to synthesize, review, or update project files from it, with authenticated recovery, scope-preserving acceptance tracking, explicit completeness boundaries, bounded independent mathematical repair, durable deferred-task recording, and downstream validation.