Agent Skill Discovery

Agent Skill Discovery

Agent Skill Discovery is the workflow of finding, ranking, reading, and selectively adapting reusable agent skills before installing or copying them. It treats public skill hubs as a research corpus, not as a package manager to trust blindly.

Workflow

1. Discover from multiple indexes: Hermes Skills Hub, skills.sh, GitHub repositories, and official platform skill collections.
2. Rank by both popularity and trust: install count/star count, official source, repository reputation, recency, and specificity.
3. Archive raw content before interpretation so future sessions can re-read the original instructions.
4. Read for patterns: trigger design, precondition checks, tool commands, fallback paths, verification gates, and “do not use for” boundaries.
5. Synthesize into wiki pages before installation.
6. Install or adapt later only when a specific recurring task justifies it.

Quality signals

• Concrete trigger phrases in description:.
• Clear scope boundaries and negative triggers.
• Stepwise workflow with prerequisites before action.
• Verification steps and observable success criteria.
• References/scripts separated via progressive disclosure.
• Security boundaries for credentials, writes, purchases, deployment, or browser automation.

Anti-patterns

• Installing high-install skills without reading them.
• Treating install count as equivalent to quality.
• Copying commands for tools Hermes does not have.
• Letting a script own agent orchestration instead of keeping reasoning in the main agent.

Related

• Agent-Skills-Hub-and-Skills-sh-High-Star-Skills
• AI-Self-Improvement-Lab

2026-06-30 补充：从技能文件到技能分发基础设施

Agent Skills 生态快照 2026-06-30（agent-skill-ecosystem-2026-06-30）显示，外部技能生态正在从零散 SKILL.md 发展为分发基础设施：marketplace、package manager、lockfile、curated directory、cross-agent compatibility。代表项目包括 craftdesk、cc-skills 和 skill-of-skills。

因此 discovery workflow 的重点应继续保持在“先读再装”：星数和 marketplace 只能作为发现信号，不能替代对触发条件、权限边界、验证步骤和副作用的审查。

2026-08-30 补充：发现 skill 之后还要审计派生 skill

EVOMAL（evomal-self-poisoning-agent-skill-libraries）把 Agent Skill Discovery 的安全边界从“外部 skill 是否可信”扩展到“agent 模仿外部 skill 后生成的派生 skill 是否可信”。共享 skill library 中的恶意模板即使未被直接调用，也可能在 authoring 阶段被复制进新 skill，并通过 agent-authored copies 留在长期库中。

因此 discovery workflow 应新增派生审计：任何从外部 README、网页、skill hub 或历史 skill 生成的新 skill，都应记录 derived_from、source_hash、author_run_id、review_status，并在进入自动加载前检查动态下载/执行、凭据读取、网络外传、混淆 payload、过宽权限和 banner-style copy。

2026-10-24 补充：恶意标签、拒绝和危害终态分开

主题簇归属： 混合技能审查归入Harness-Engineering的权限/持久控制面，以及Agent-Benchmarks的独立oracle与合法效用；正文导航不能只写轴名或依赖来源页间接连接。

MalSkillBench 方法审计（malskillbench-malicious-label-versus-harm-and-refusal；sandbox-security / benchmark-evaluation / harness-runtime）补充三个检查：联合读取合法任务意图、技能指令及代码/配置；把 hooks、身份/目标和未来 session 载入视为持久控制面；分开恶意标签有效、识别/拒绝与实际危害。该论文 PI 样本的 detected 也可作 positive，且两层证据采用“或门”，不能把 runtime-verified 全翻译成成功攻击。

发现仍只授权只读审查与知识入库，不自动安装；若另获隔离测试授权，使用合成凭据/可控接收端、合法正对照、独立 effect 读回与可信恢复，并按 campaign/来源分层评测。本文未扫描或修改用户生产技能，也未把摘要与细分表格不一致的 detector 榜单当采购依据。

写入记录

• 2026-08-30 09:00 CST：补充 EVOMAL 对 skill discovery、派生 skill provenance 与安全审计字段的启发。
• 2026-10-24 09:10 CST：补 MalSkillBench 的代码/指令/持久控制面联合审查、标签/拒绝/真实危害分账及主题反链；不安装技能或执行恶意样本。
• 2026-10-24 09:15 CST：把生态快照的Obsidian alias链接改为已核实的文件stem加显示名，兼容现有lint/静态发布器，不改目标文章或历史结论。
• 2026-10-24 12:04 CST：汇总巡检发现只有主题轴文字、没有六轴主题概念正文直链，补Harness/Benchmark两处入口；不改原研究结论、技能文件或生产权限。
• 2026-10-24 12:09 CST：精确HTML验收发现三处带显示名wikilink的竖线被现有发布器误认成表格；改为显式stem的普通wikilink并保留相邻中文/项目名，修复生态快照、EVOMAL和MalSkillBench真实href，不改发布器或批量迁移。