Skills Ecosystem

What's in this pack

Skills are the format Anthropic standardized in late 2025 for distributing reusable agent capabilities — a folder containing a SKILL.md plus optional scripts, references, and templates. Vercel and others adopted the same standard. This pack collects the seven assets that let you write a skill once and ship it across the entire AI tooling ecosystem.

Why this matters

Before the skill standard, every AI tool wanted the same capability described in its own dialect: a Cursor rule was an .mdc, a Claude Code subagent was a Markdown frontmatter file, a Codex CLI project wanted AGENTS.md, a Cline rule lived in .clinerules. The same logical skill (e.g. "review code for security issues") had to be hand-translated 14 times.

The skill standard solves this with one canonical format and a tooling layer that emits the per-tool variants. Three concrete wins:

1. Single source of truth. You author and version-control one SKILL.md per capability. Reviews, edits, and changelog all happen in one place.
2. Mechanical translation, not copy-paste. The converter tools generate Cursor .mdc, Codex AGENTS.md, Cline rules, and the others from the canonical skill. No drift between platforms.
3. Composability at the agent level. Anthropic's runtime can load multiple skills into one Claude session. Vercel's AI SDK implementation does the same across model providers. A skill written today still works in two years when the underlying model changes.

Install in one command

# Install the entire pack
tokrepo install pack/skills-ecosystem

# Or pick the piece you need first
tokrepo install skill-maker
tokrepo install vercel-skills

The TokRepo CLI drops skill scaffolding, the converter scripts, and a starter SKILL.md template into your project. Run one command and you have a skill that builds for all 14 targets.

Common pitfalls

• Treating SKILL.md as a prompt template. It's not — it's a capability declaration with metadata, allowed tools, and progressive-disclosure references. Write it as you'd write an internal documentation page, not as an LLM prompt.
• Including secrets in the skill body. Skills are designed to be checked into a public repo. Reference secrets via env vars or runtime injection; never paste them in.
• Skipping the description field. The agent uses the description to decide whether to load this skill on a given query. A vague description means the skill never gets used. Be specific about scope and triggers.
• Forgetting the per-target feature matrix. Cursor .mdc supports glob-based scoping. Codex AGENTS.md is project-wide. Claude Code subagents have allowed-tools. The converters emit the closest equivalent, but you have to be aware of the differences when you author.
• Building a skill for one tool when you mean to build it for all. If the capability is portable (review code, summarize a document, generate tests), author as a skill from day one. If it's tool-specific (a Claude Code plugin that uses Anthropic's tool API directly), then write it as a tool-specific asset.

Relationship to other packs

The Anthropic Builders pack ships the 17 first-party Anthropic skills that establish the canonical patterns. Use those as references when authoring your own. The Awesome Claude Code Subagents pack is the runtime layer that loads skills into a Claude Code session — it's what runs your skill once it's authored. This Skills Ecosystem pack focuses on the cross-platform translation layer: how to take a skill and make it work in Cursor, Codex CLI, and the rest.

For tools that don't yet support a skill format natively, the converter outputs include a "shim" asset (a Markdown file the user pastes into the tool's context) so even unsupported surfaces get most of the value.