Skip to main content

Router (Reference)

What this is​

.cursor/prompts/20-router.md is the Router prompt. It tells the AI how to choose between SPEC and IMPLEMENT behavior using the Context Pack (platform type, exposure, data sensitivity) and to ask one question if platform type is missing.

When to use it​

  • When you need the AI to switch or stick to SPEC vs IMPLEMENT based on the Context Pack.
  • When the AI is mixing planning and code in one response; paste the Router to enforce the split.

Steps​

  1. Ensure your Context Pack is in the chat (platform type, exposure, data sensitivity).
  2. Paste .cursor/prompts/20-router.md (or the exact text below) when you want SPEC/IMPLEMENT routing applied.

Exact text (from kit)​

Source: .cursor/prompts/20-router.md

# Router

Use this when you need the AI to choose between SPEC and IMPLEMENT behavior based on the Context Pack.

---

**Inputs (from Context Pack):**
* Platform type: the slug from `docs/ai/ai-config.md` (e.g. `consumer-app`, `data-platform`, `saas-b2b`)
* Exposure level: [ internal | external-authenticated | public ]
* Data sensitivity: [ public | internal | confidential | restricted ]
* Current task or question

**Rules:**
* If platform type is missing and cannot be inferred from context: ask **exactly one** question (e.g. "What platform type should I use? Check `docs/ai/ai-config.md` for the slug.") and stop.
* **SPEC behavior:** User said "Switch: SPEC" or the ask is clearly planning/scoping. Output: scope, acceptance criteria, open questions. No code. Apply PM / Discovery / Design / Security agents as relevant; use exposure and data-sensitivity overlays when security is in scope.
* **IMPLEMENT behavior:** User said "Switch: IMPLEMENT" or the ask is clearly code/build. Output: files to change, plan, diffs, verification. Proceed with minimal safe changes. Apply FE / QA / Validation / Security agents as relevant; use exposure and data-sensitivity overlays when security is in scope.

**Output:** Proceed in the chosen mode. Do not mix SPEC and IMPLEMENT in one response unless the user explicitly asks for both.