跪拜 Guibai
← All articles
Artificial Intelligence · AI Programming

AI Agent Skills Explained: The Folder Standard That Cuts Context Bloat

By oil欧哟 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Skills cut token overhead and make agent workflows portable across tools, since the open spec is now supported by Codex, Cursor, and others. The warning against installing too many vague Skills is a practical safeguard that keeps agents focused.

Summary

Anthropic's open-source Agent Skills specification has become a cross-tool standard. A Skill is simply a folder with a SKILL.md entry and optional scripts, references, and assets. Its progressive loading keeps only each Skill's name and description in context, loading the body only when triggered. That slashes token costs and avoids scattering the model's attention.

Skills split into knowledge-based (encoding judgment like style rules) and process-based (fixing multi-step workflows with scripts). The guide recommends using global rules for always-on conventions, quick prompts for one-liners, and Skills for specific know-how or scripted processes.

Installing is easy: manual copy, npx skills add, or asking the Agent itself. Trigger via automatic matching or manual slash commands. The core warning: don't install Skills before you hit a real problem, and avoid vague AI-generated Skills because they degrade performance and too many installed Skills confuse the Agent.

Takeaways
— Anthropic open-sourced Agent Skills as a folder-based specification; Codex, Cursor, and DeepSeek Harness now support the same structure.
— A Skill folder requires a SKILL.md entry with YAML frontmatter containing name and description; optional scripts, references, and assets directories hold supporting files.
— Only the name and description of every installed Skill stay in context; the body and attachments load only when triggered, cutting token use and attention loss.
— Skills fall into two types: knowledge-based (like oil-tone, which encodes writing judgment) and process-based (like Oil UI, which encodes multi-step workflows with scripts).
— Global rules (AGENTS.md/CLAUDE.md) belong to always-on conventions; quick prompts suit one- or two-sentence reminders; Skills fit specific knowhow or multi-step scripts.
— Install Skills by manual download, via npx skills add owner/repo, or by asking the Agent itself to download and place them.
— Trigger Skills automatically when the Agent matches the description, or manually by naming the Skill or using a slash symbol.
— Avoid installing Skills before hitting a real problem; vague, AI-generated Skills hurt performance, and too many installed Skills scatter the model's attention.
Conclusions

Rapid adoption of Anthropic's Skill spec across competing tools shows that a plain folder convention becomes a standard when it solves a universal pain: context budget.

Separating knowledge from process is a useful mental model for deciding when to invest in a Skill versus a global rule or a quick prompt.

The warning against installing many Skills runs counter to the 'more is better' instinct, but the constant presence of name/description metadata makes every added Skill a potential source of confusion.

Concepts & terms
Agent Skill
A folder containing a SKILL.md file and optional scripts/references/assets, with a defined structure that any AI agent can read to load task instructions on demand.
Progressive loading
A mechanism where only a Skill's name and description are kept in the context window; the full SKILL.md body and attachments load only when the agent matches the task.
Knowledge-based Skill
A Skill that encodes personal experience, aesthetics, and judgment criteria to correct a model's default behavior, without necessarily adding scripts.
Process-based Skill
A Skill that solidifies a multi-step workflow, often including executable scripts, so an agent can run the entire process consistently.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗