Stim

Stim

**A DSL for writing AI prompts — commands, agents, and rules — that compiles to whatever tool you use.**

Write one `.stim` file. Target Claude Code, ChatGPT, Cursor, or any future AI tool by switching a flag.

[](https://opensource.org/licenses/MIT)

---

Why Stim?

AI prompts have become software. Slash commands, system prompts, agent definitions, rule files — they're all code now, but most of us write them as raw markdown that breaks the moment the logic gets complex or the tool changes. Stim treats prompts as source code: parse, validate, compile.

**Three declarators, three targets:**

command deploy {
  ask("Which environment?")
  wait_for_response()
  for service in services {
    task bash "deploy " + service { ... }
  }
}

agent reviewer {
  description "Reviews PRs for security and correctness"
  tools [Read, Grep, Bash]
  model "sonnet"

  "You are a senior code reviewer. Cite specific files and line numbers."
  "Prioritize by severity: must change, should change, could improve."
}

One source, many outputs:

stim install reviewer.stim                   # → ~/.claude/agents/reviewer.md
stim install reviewer.stim --target cursor   # → .cursor/rules/reviewer.mdc
stim compile reviewer.stim --target chatgpt  # → dist/chatgpt/reviewer.md

What You Get

• Three declarators — command for interactive slash commands, agent for system prompts with metadata, harness for reusable orchestration templates
• Three targets — claude, chatgpt, cursor (more to come; targets are pluggable adapters)
• Real control flow — if/else, while, for, break
• Orchestration — bounded agentic loops, dynamic fanout/gather, content routeing, and parameterized harnesses you run
• Variables and imports — strings, booleans, arrays; share definitions across files
• Subagent tasks — spawn bash, explore, plan, or `gene