Slash commands → Skills: what changed and why¶
The core shift: Slash commands required explicit invocation by the user. Skills are loaded automatically — Claude scans available skills and pulls in the right one based on task context. No
/commandsyntax needed.
Agent Skills launched October 2025 and replaced custom slash commands as the primary extensibility model for Claude Code plugins and OpenAI Codex plugins.
What Skills are¶
A Skill is a folder containing a SKILL.md file plus any supporting scripts and resources. Claude loads only the minimal content from a relevant skill when needed, keeping response quality high without bloating context.
| Property | Description |
|---|---|
| Composable | Multiple skills stack automatically. Claude identifies which are needed and coordinates them. |
| Portable | Same format everywhere — Claude apps, Claude Code, and the API all use the same skill structure. |
| Efficient | Only loads what's needed, when it's needed. Context is not polluted with unused instructions. |
| Executable | Skills can include runnable code for tasks where traditional programming is more reliable than token generation. |
Two kinds of skills¶
Understanding which type you're building affects how you write and test it.
Type 1: Capability uplift¶
Helps Claude do something it can't do consistently on its own — e.g. the document-creation skills (docx, pptx, pdf). These may become less necessary as base models improve. Evals can tell you when that's happened and the skill is no longer needed.
Type 2: Encoded preference¶
Claude can already do each individual step; the skill sequences them according to your team's specific workflow — e.g. NDA review against set criteria, or generating CMS block patterns from a Figma design. These are more durable over time. Evals verify fidelity to the actual workflow.
Installing skills¶
Claude Code¶
Skills in Claude Code are installed in one of two ways:
- Via plugins from the
anthropics/skillsmarketplace - Manually by placing skill folders in
~/.claude/skills
Claude loads them automatically when relevant. Skills can be shared across a team through version control — just commit the skill folder.
OpenAI Codex¶
Skills in Codex are distributed as part of plugins. Install a plugin via the Codex plugin browser (codex/plugins) or CLI (codex plugin marketplace add owner/repo), and its bundled skills become available automatically. Codex loads skills from the skills/ directory specified in .codex-plugin/plugin.json.
Skills activate by natural language context, or invoke them explicitly with @skill-name in the prompt.
Plugin structure: the container for skills and agents¶
The plugin is the distribution and installation unit — it's what gets installed via claude plugins install and what lives in version control. Skills and agents are the functional components that live inside it.
| Layer | Role |
|---|---|
| Plugin | Container/package — handles installation, versioning, and scoping (project-level or global) |
| Skills | Encoded workflow knowledge — SKILL.md files that Claude pulls in automatically when context matches |
| Agents | Autonomous, goal-directed task runners — can use tools and execute multi-step workflows independently |
This means a plugin like cms-cultivator is the right home for both skills and agents. They serve different purposes within the same package:
- Skills handle context-triggered, Claude-assisted tasks — e.g. generating a WordPress block pattern from a Figma design, or scaffolding a Drupal paragraph type from a screenshot. Claude loads these automatically when it recognizes the task.
- Agents handle heavier workflows that need to run autonomously — e.g. auditing all blocks in a theme for accessibility, or migrating templates to FSE format. A subagent is spun up with a clear goal and the tools it needs to complete it.
One important implication: skills distributed inside a plugin use the same description-matching mechanism as standalone skills. The description precision work that skill-creator helps with (see below) applies equally to plugin-bundled skills. Running the description optimizer across plugin skills is worthwhile after any significant edits or after a model update.
Testing and maintaining skills with skill-creator¶
Skill-creator (available as a Claude Code plugin and for OpenAI Codex) adds software-development rigor to skill authoring — no code required.
Evals¶
Define test prompts, describe what good output looks like, and skill-creator checks whether the skill holds up. Two primary uses:
- Catching regressions after model updates
- Knowing when base model capabilities have outgrown a capability-uplift skill (it passes evals without the skill loaded)
Benchmark mode¶
Runs a standardized assessment tracking eval pass rate, elapsed time, and token usage. Run it after model updates or after editing a skill. Results can be stored locally, piped to a dashboard, or integrated with CI.
Multi-agent eval runs¶
Evals run in parallel using independent agents — each in a clean context with its own metrics. Comparator agents support A/B testing of two skill versions (or skill vs. no skill), judging outputs without knowing which is which.
Description optimization¶
Skill-creator analyzes your skill description against sample prompts and suggests edits that reduce both false positives and false negatives. This is the key lever for reliable triggering — too broad and unrelated tasks load the skill, too narrow and it never fires.
Looking ahead¶
Today a SKILL.md tells Claude how to do something. As models improve, a plain description of what the skill should do may be sufficient — with the model working out the rest. The eval framework is a step in that direction: evals already describe the "what."
Official docs and resources¶
Blog posts¶
- Introducing Agent Skills (Oct 2025) — The launch post covering the full architecture, how skills work across all Claude products, and getting started links.
- Improving skill-creator: Test, measure, and refine Agent Skills (Mar 2026) — Covers evals, benchmark mode, multi-agent testing, and description optimization.
- Improving frontend design through Skills (Nov 2025) — Real-world example of a capability-uplift skill improving Claude's design output quality.
- Engineering blog: Agent Skills architecture deep-dive — Design patterns, architecture, and development best practices from the Anthropic engineering team.
Documentation¶
- Claude Code Skills docs — Installing, building, and sharing skills specifically within Claude Code.
- Agent Skills overview (API docs) — Full reference covering the Skills format, the
/v1/skillsendpoint, and Code Execution Tool requirements. - User guide: Teach Claude your workflow using Skills — End-user guide for creating and managing skills in Claude apps.
GitHub repos¶
- anthropics/skills — Official skills repo — Anthropic's public skills including
docx,pptx,pdf,skill-creator, and examples to customize. - skill-creator plugin for Claude Code — Install this plugin to get skill authoring, evals, benchmarking, and description optimization directly in Claude Code.
- skill-creator skill source — The skill-creator skill itself — useful for understanding how Anthropic structures their own skills.