Scaffold a new resource

How the grill workflow forces a structured Q&A before any file is written.

The fastest way to produce a non-trivial new FiveM resource through an agent is to invoke the grill workflow. It forces the agent to ask every architectural question before any file is written.

Triggering the workflow

> /agent_api:scaffold-fivem-resource
> สร้าง resource ใหม่ที่...
> create a new fivem resource for ...
> build a script that ...

The intent-detection path works in every client. The slash-command path is Claude Code only (other clients don't surface MCP prompts).

What the agent will do

It cannot skip steps — its mandatory contract is loaded into context.

Step 0 — preferences first. Before the tree, the agent calls list_preferences. For every saved preference (structure / coding / ui-design) it asks, one at a time, whether to adopt it as the guideline for this resource — and biases the rest of its recommendations toward the ones you accept. Set these up in the dashboard Preferences tab.

The agent also reviews the conversation and skips questions you've already answered (confirming instead of re-asking). Question tree in order:

  1. Identity — name, description, author
  2. Layout — one file vs multi-file with server/, client/, shared/, config/ subfolders. Flat layouts forbidden.
  3. Module style — globals via manifest vs return {} + lib.require (needs ox_lib)
  4. Config split — config/shared.lua, config/client.lua, config/server.lua. The last one stays server-side and is the only safe place for secrets.
  5. Framework — standalone, ESX, QBCore, ox_core, multiple
  6. ox_lib — yes / no
  7. oxmysql — none / yes (+ list of tables to own)
  8. UI — none / yes
  9. UI stack — pure HTML+JS / Vite framework
  10. Framework choice — Vue 3 / React / Svelte / Solid / Preact
  11. State management — built-in / Pinia/Zustand/etc / none
  12. CSS approach — plain / scoped / Tailwind / UnoCSS
  13. Animations — none / CSS transitions / framework transitions / GSAP
  14. Validation — none / hand-rolled / Zod/Valibot
  15. Icons — none / inline SVG / Lucide / Phosphor
  16. Open key / command — F6 + /<name> recommended
  17. NUI focus model — modal / overlay / hybrid
  18. Lifecycle — start, playerDropped, stop behaviour
  19. Audit + rate limit — what to log, what to throttle

Every question carries lettered options (A / B / C / D …) and a "Recommended:" line, so you can reply with just a letter or "ok" / "เอา" and move on. Before it builds, the agent replays everything as a numbered summary; to revise a decision, say edit <number> referring to that line.

Then it builds

Once you confirm the summary, the agent walks the scaffolding phase:

  1. create_resource
  2. refresh_resources
  3. If UI = Vite: run_shell npm create vite ..., run_shell npm install, edit vite.config, run_shell npm run build
  4. write_file every Lua / config file (sandbox-checked)
  5. ensure_resource
  6. If state didn't reach started: tail_console, propose a fix, ask before rewriting

You see every step. Nothing happens silently.

PreviousWeb dashboardNextLive UI testing