Quick answer: in 2026 agent configuration split into three layers — ambient AGENTS.md repo rules, invokable SKILL.md procedural workflows (Linux Foundation Agentic AI Foundation standard, zero ambient context cost), and MCP servers for live tool access. This site converts your legacy .cursorrules or system prompt into both formats, client-side, for free.
.cursor/rules/*.mdc and AGENTS.md replaced the single legacy .cursorrules file. AGENTS.md is now used by 60,000+ open-source repos and is the cross-agent standard read by Claude Code, Cursor, Codex, Windsurf and Continue; SKILL.md carries YAML frontmatter and loads only when a task matches.
| Layer | 2026 Format | Context Cost |
|---|---|---|
| Ambient repo rules | AGENTS.md + .cursor/rules/*.mdc | Always loaded |
| Invokable workflows | SKILL.md (YAML frontmatter + 100–400 line body) | Zero until invoked |
| Live external tools | MCP (Model Context Protocol) | Per-tool schema |
Paste your legacy rules below. The compiler separates persistent repo conventions (→ AGENTS.md) from repeatable procedures (→ SKILL.md with spec frontmatter), then scores the skill against the 100–400 line progressive-disclosure window. Everything runs in your browser — nothing is uploaded.
—
—
Three ways: drop the file in .agents/skills/<name>/SKILL.md (project) or ~/.agents/skills/ (global); use the community installer npx skills add <repo>/skills@<name>; or paste the raw markdown URL into Claude Code / Cursor as project instructions.
Keep the reasoning LLM on planning and prose; route closed-set classification, safety screening and routing gates to a single-forward-pass decision schema. A SKILL.md for a JEV twin workflow should describe when to call the decision endpoint, never re-implement the classifier in prompts. See the companion spec hub at decisionprompts.com.
Raw /skills/<slug>/SKILL.md endpoints ship with the full library — this page is the interactive entry point.