Projects, Custom Instructions, and Custom Assistants
Set up chat assistants so they already know your context: projects, custom instructions, knowledge files, and shareable custom assistants.
If you use Claude, ChatGPT, or Gemini every day, you probably start a lot of conversations the same way: "I'm a product manager at a 40-person startup, we sell to dental clinics, keep it short, here's our style guide again." Every new chat starts from zero, so you rebuild the same context by hand, and on the days you skip it, the answers get generic.
All three major chat apps let you fix this permanently. You can give the assistant standing instructions about you, set up workspaces that hold the files and rules for a specific stream of work, and package an assistant for a recurring job so you, or your whole team, can reuse it with one click. This lesson shows you how each layer works, what to put in it, and how to keep it from going stale, so every conversation starts with the right context already loaded.
Three layers of configuration
Chat apps name these features differently, and the details change often, but they all offer the same three layers. Think of them as nested scopes, from broadest to narrowest:
| Layer | Scope | Typical names | Best for |
|---|---|---|---|
| Personal instructions | Every conversation you have | Custom instructions, personal preferences, saved info | Who you are and how you like answers |
| Projects | Conversations inside one workspace | Projects (Claude, ChatGPT), notebooks and workspaces in other tools | An ongoing stream of work with its own files and rules |
| Custom assistants | Anyone who opens that assistant | Custom GPTs (ChatGPT), Gems (Gemini), shared projects (Claude on team plans) | A repeatable job you or others do often |
All three are just context engineering with a friendly interface. Whatever you type into these settings ends up in the model's context, much like a system prompt (see Writing System Prompts That Hold Up). The same principles apply: be specific, explain why, and keep only what earns its place.
There's a fourth thing that looks similar but works differently: memory, where the app itself decides what to remember from your chats. That's covered later in this module, in Giving Agents Memory. The difference matters: memory is what the assistant learned; the layers in this lesson are what you decided. You control them completely, so they're where your most important context belongs.