One honest job: structure Grok completely. Seven layers, same mission as August, current product underneath. Then a first setup you can finish in one sitting.
01 · Foundation
Global Custom Instructions
Settings → Custom Instructions. Voice, core principles, default format, and how uncertainty is labeled. Keep this layer light. Domain weight lives in a Project. Plain English. No private names in anything you might share.
02 · Personas
Custom Agents
An agent is a person you can return to, not a task. Give each one a name, a narrow instruction, and only the tools that role needs. How many slots your plan allows is not proven from the public plans page — open Agents and count them. On Heavy, the plans page says a larger team of agents collaborates. Do not invent a headcount.
03 · Domains
Projects / Workspaces
One Project per major area of work. Instructions, files, and history stay with that domain. Global stays modular. Do not pour every file into one drawer.
04 · Continuity
Memory — two stores
Chat memory keeps preferences and past conversations across chats. Build memory is different: project notes, written after a turn, read back later, browsable with /memory and merged with /dream. Curate both. Delete what is stale. Never paste either store into a public file.
05 · Repeatable work
Skills
A skill is a task, not a persona. Teach it once. It can apply across conversations. Built-ins shipped for documents, decks, spreadsheets, PDFs, and a skill creator — confirm the live list, because it can grow. Your skill overrides a built-in when both could apply. Every example in this guide is replaceable.
06 · Force
Modes and tools
Expert mode for a harder answer. Heavy, on the Heavy plan, for a larger team of agents on one hard problem. Imagine, Voice, and live search when the job needs them. Build when the sitting is a ship — an app, a page, a tool — not when the sitting is talk. Do not turn every tool on by default.
07 · Your voice
Personalization
After real use, tighten the rules. Preferred phrasing, the words in your domain, and what happens when you say X. Structure is living. Five to ten honest sittings beat a perfect first draft.
First setup
- 01Write Global Custom Instructions: voice, honesty labels, one sitting, human decision.
- 02Open Plans in your account. Do not trust an August screenshot.
- 03Create agents only for roles you actually repeat. Leave the rest blank.
- 04Open one Project for the highest-priority domain. Put files there, not in Global.
- 05Turn memory on. Prune it. Treat chat memory and Build memory as two drawers.
- 06Write one skill for one repeated job, or override one built-in. Not fifty on day one.
- 07Run one real sitting. If it is talk, do not build.
- 08After five to ten uses, correct the layer that failed. Then say what the next pass should prove.