Make an AI Assistant: Give It Rules, Context, and Boundaries

Making an AI assistant that works well for your job means more than picking a chat app. You need to give it a focused task, write instructions it can actually follow, and decide what it may and may not do. An assistant without clear boundaries either refuses too much or goes too far.

Persistent context is the other half. A general-purpose chat window starts from zero every session. If you connect your assistant to a tool like fenbs, it can read your current projects, rules, and where things were left before it acts, and save progress when it stops.

Choose a single job first

The most common mistake is trying to make one assistant do everything. An assistant configured to "help with all my work" has no clear standard of correctness. An assistant configured to "draft and review client email responses in the tone of our brand guide" can be tested, iterated on, and trusted for that specific task.

Pick one recurring job where you currently spend time that the assistant could do well. Good candidates: summarizing meeting notes, drafting first versions of documents you write repeatedly, answering questions about a knowledge base you maintain, or reviewing a specific type of document against a checklist. Start narrow. Expand later.

Pick an AI app and configure its instructions

Most AI apps let you provide instructions that the model reads before every conversation. In Claude, these go in a Project. In ChatGPT, they go in the system prompt or a custom GPT. In Cursor and other coding tools, they go in a configuration file like AGENTS.md.

Write the instructions as if briefing a capable new colleague. Tell them: what you are working on, the standards to follow, the format to use, what to ask about rather than decide, and what to do when they are not sure. Short, direct instructions work better than long paragraphs.

For the range of AI apps available for different jobs, AI desktop assistants and AI assistants for developers cover specific categories.

Define the assistant's access and limits

Access means what the assistant can read and change. Limits mean what it may not do without asking. Both matter.

A well-scoped assistant for email drafting can read your previous emails for style context, but does not send email without your review. An assistant that helps manage a project board can update task status, but does not close tasks or change priorities without asking. Write those limits in the instructions explicitly.

• Which tools or files the assistant may read.
• Which actions it may take without asking.
• Which actions always need your review before they happen.
• What it should do when it encounters something unclear.
• Which topics it should not address.

Test its boundaries before trusting it with real work

Before you rely on an assistant for live tasks, test it. Ask it questions at the edges of its scope. Ask it to do something it should not do. Ask it an ambiguous question and see whether it asks for clarification or guesses. A bad answer in testing is useful information. A bad answer in production is a problem.

Expect to revise the instructions. Most instruction sets improve significantly after the first round of testing. If the assistant is getting something consistently wrong, the instructions are probably unclear about that case.

Share context across sessions

A chat session ends and the context is gone. This is the main limit of most AI apps for ongoing work. You rebuild the context at the start of every session, which takes time and introduces inconsistency if you describe the project slightly differently each time.

Storing context outside the chat solves this. The context might be a document, a project file, a knowledge base, or a connected tool like fenbs. The assistant reads it at the start of the session. The job is to keep that external context current and accurate, not to repeat it in every chat.

A shared external record also works when you switch apps. If you switch from one AI assistant to another, the new assistant reads the same context from the same place. The instructions your assistant follows do not depend on which specific app you are using today.

How to do this with fenbs

1. Create a free account. Add your project: what it is, the current tasks, and the rules that apply.
2. Write the rules your assistant should always follow as Decisions and rules. These become the first thing a connected assistant reads in each session.
3. Connect your AI app from Integrations. Supported apps include Claude, Claude Code, ChatGPT, Cursor, Codex CLI, Gemini CLI, GitHub Copilot, and Devin Desktop.
4. At the start of each session, ask the assistant to read the project context before working. At the end, ask it to save Where we left off and update the task statuses.
5. When something important changes, update the relevant Decision or rule. The assistant reads the current version, not the version from when you first configured it.

fenbs does not configure the AI app for you. It gives the assistant the context it needs to work well on your project across multiple sessions and apps. For the broader how-to on building a custom assistant, custom AI assistant covers platform-specific options.