Standing Setups: Assistants That Remember¶
Getting Better Answers explains the three levers that decide output quality: context, memory, and standing instructions. This page is about making the third lever permanent. If you find yourself re-explaining your course, your project, or your preferences at the start of every conversation, you are doing setup work that the tools are designed to hold for you. Set it up once, and every future session starts already knowing your job.
The Four-Minute Version
The Three Containers¶
| Container | Where it lives | What persists | Best for |
|---|---|---|---|
| Claude Project | claude.ai | Per-project instructions, an uploaded knowledge base, and a separate per-project memory | A course, a study, a committee: any body of work with stable materials |
| ChatGPT Project | ChatGPT app and web | Project instructions, uploaded sources, and the project's chats | The same jobs, on the OpenAI side |
| Folder brief | A markdown file in a folder an agent works in (CLAUDE.md or AGENTS.md, depending on the agent) |
Standing rules the agent reads before any work in that folder | Recurring agent work on the same files |
The common idea: instructions plus materials, attached to the work instead of the conversation. A project's instructions apply to every chat inside it, and its uploaded knowledge is available without re-uploading. Each company page's features table shows what its assistant keeps between conversations: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), and Grok (xAI). Google's version is the notebook: a per-course Gemini Notebook is the same pattern for source-grounded questions, your lectures uploaded once and answered from with citations all term.
There is a fourth container worth knowing about, and it works the other way around: a skill attaches instructions to a kind of task rather than to one body of work, so it applies wherever you are. If you use claude.ai or Cowork, four of them are already available to you, and that page also covers how to judge a skill someone else wrote.
Anatomy of a standing setup
The container: a project, or a folder with a brief
Set up once; every conversation inside inherits both boxes above it.
How memory works inside a project differs from one company to the next, but one rule holds everywhere: anything that must always apply belongs in the project's instructions, not in its memory. OpenAI makes the point in its guidance for Codex: "Treat memories as a helpful recall layer, not as the only source for rules that must always apply."
Worked Pattern: A Course Assistant¶
The highest-value standing setup for faculty is one project per course:
- Create the project and name it for the course.
- Upload the knowledge: syllabus, learning objectives, the session schedule, your reading list, and any handouts you would give a student. A course assistant has no use for rosters, grades, individual student work, or identifiable patient details, and leaving them out keeps the project easy to share with a colleague or a teaching assistant.
- Write the instructions, which are just a standing version of a good prompt: who the assistant is (a teaching assistant for this specific course and level), what it answers from (the uploaded materials first, citing the section it drew on), and your standing rules (US English, the course's terminology, "say so when the materials do not cover a question rather than guessing").
- Use it all term. Draft announcements, generate practice questions against the actual objectives, and check whether a planned session duplicates an earlier one, all without re-explaining the course once.
Students can run the same pattern in reverse: a project per course, lecture notes and syllabus as knowledge, and the National Board of Medical Examiners (NBME)-style question tutor from the prompt library pasted in as the project instructions. That combination, a reviewed prompt plus your own materials in a standing container, is the study setup this site recommends.
The same shape serves research: one project per study, with the published papers you have appraised as its knowledge (the literature review playbook's synthesis step, made permanent) and the project instructions carrying your reporting guideline and journal target.
Folder Briefs: Standing Instructions for Agents¶
Agent interfaces have their own version, and it is just a text file the agent reads before doing any work in a folder. Claude Code reads CLAUDE.md; Codex, Google's Antigravity, and xAI's Grok Build read AGENTS.md, and Claude Code reads an AGENTS.md too when a folder has no CLAUDE.md, so one brief can serve several agents. Each agent also reads a personal brief from your home folder that applies to all your work; the company pages give the details. What belongs in a brief is what you would otherwise repeat: where things are, what conventions to follow, the "always do X" rules. Two practical notes: Claude Code and Codex can each draft a starting brief for you (the /init command, with worked examples on the Claude and ChatGPT pages), and the brief should stay short, since Anthropic's guidance recommends under 200 lines; a brief the length of a policy manual stops being read carefully, by models as by people. Some agents also keep their own notes of your corrections and preferences (Claude Code's /memory shows them); glance at them now and then, since they shape every later session.
The heuristic for what to add: when the agent makes the same mistake twice, or you type the same correction twice, that correction belongs in the brief. Standing setups are how one-time feedback becomes permanent behavior.
How folder briefs layer
Your personal brief
in your home folder: the defaults that apply to all your work
The folder's brief
CLAUDE.md or AGENTS.md: this project's rules, on top
The session
the agent reads both briefs, global to folder, before any work
The same correction twice becomes a line in the folder's brief.
The personal brief carries your voice everywhere; the folder's brief speaks for the project.
Keeping a Setup Accurate¶
- Instructions are advice to the model, not enforcement. A project instructed to answer only from its knowledge will still occasionally reach beyond it; spot-check citations against the uploaded materials, especially early on.
- Projects accumulate. Review a long-lived project's knowledge each term; a stale syllabus in the knowledge base produces confidently outdated answers, which is worse than no assistant at all.
Next: Agent Skills, the last step of the pathway's agent stage.
Content last reviewed October 8, 2026. Review dates are maintained as described on the About page.