Onboard your
AI.
A new AI assistant should not need to guess who you are, how you work, or what it is allowed to touch. Start with a short working brief and a test using fictional information.
Start with one task and no connections.
An LLM can be remarkably capable, but it has no lived understanding of your preferences on day one. It does not know your audience, your history, or which decisions carry consequences. Your job is to give it enough of the right context to be useful - without handing it everything.
Good onboarding has four parts: set the environment, conduct a calibration interview, choose context carefully, and test the behavior. You do not need email or calendar connections for the first exercise.
Environment and guardrails
Choose what it can remember, what it can access, and what always needs your approval.
Calibration interview
Let the assistant ask the high-value questions that reveal how you work and communicate.
Scoped context feeds
Connect useful sources deliberately instead of flooding the model with your entire digital life.
Operational test run
Give it realistic ambiguity, formatting, and edge cases before relying on it for important work.
Phase 1: set the environment
Start with permissions before personality. Decide what the assistant can read, what it can change, and which actions require you to say yes. Default to the smallest access that allows the task.
Useful, not automatic
If a platform offers persistent memory, use it only for stable facts you are comfortable retaining: your preferred communication style, your role, or durable constraints. Review it periodically and remove stale or overly personal details.
Match the setting to the work
Before using proprietary material, private correspondence, student information, or sensitive business plans, check the platform's data controls and your organization's policy. Do not assume every tool handles data the same way.
Read first. Confirm on write.
Use the app’s permission and approval settings where available, and limit the access you grant. Instructions to ask first are useful, but a prompt alone does not enforce a permission boundary. Test the controls before relying on them.
Phase 2: let it interview you
Do not try to write a massive autobiography. Ask the assistant to run a focused intake interview, then turn the answers into a reusable user profile or set of custom instructions that you can review.
"You are helping me set up how we work together. Ask me five practical questions, one at a time, about my background, communication style, how you should handle ambiguity, the tools and constraints in my work, and the role you should default to. After I answer, create a short user profile and operating instructions that I can review before saving anywhere."
Conversation context is what the app supplies in the current session. Saved instructions are preferences it may load in later sessions. Persistent memory is a separate, optional feature; verify what is retained and how to edit it. The best answers here are practical: Do you lead with the answer? Do you prefer a list, a table, or a narrative? When should the assistant make a reasonable assumption, and when should it stop and ask? What does helpful disagreement look like?
"Based on my answers, make a Markdown table with Rule or Constraint, Trigger Situation, and Expected AI Behavior. Then add a separate Ask Me First section: the topics, data, or assumptions you must never act on without asking first."
Phase 3: connect context with a filter
Connections to email, documents, calendars, and project systems can make an AI more useful. They can also create noise and expose more information than the task requires. A full inbox is not a knowledge base.
Ask for what you need
An app connection may still have broad account permissions even when your request is narrow; inspect the granted access separately. Ask for emails from a specific person about a specific project in the last week. Open the document that actually contains the current brief. Give a calendar view of the next few days.
Give it the whole archive
Bulk imports create clutter, make important context harder to find, and expand the surface area for sensitive information or malicious instructions hidden in external content.
Keep evergreen context in a curated project folder, workspace, or knowledge vault: the current brief, standards, decisions, and reference documents. Use live connections to retrieve a specific answer, not to swallow everything you own.
Phase 4: give it a real test run
Before an assistant becomes part of your workflow, test the behavior you care about. Do not ask whether it is smart. Ask whether it works the way you need it to work.
"I am going to give you three short scenarios. For each one, show your proposed answer, state any assumptions, and tell me whether you would proceed, ask a clarifying question, or request confirmation before acting. Keep the answer in a compact table."
“Explain the next step, then let me try.”
“I run a home business and I am new to coding. Use plain language and explain unfamiliar terms once. Give me one setup step at a time. For a reversible layout choice, make a suggestion and say what you assumed. Ask before spending money or contacting a customer. If I am overlooking a problem, tell me directly.”
Save that as editable instructions. Memory is a separate feature: it may retain preferences across conversations, depending on the app. A file called SOUL.md is simply another possible convention; it has no effect unless the tool is configured to read it.
Your first-day checklist
- My custom instructions or user profile are saved somewhere I can review and edit.
- Any persistent memory contains only durable, appropriate facts.
- I understand the platform's privacy and data-control settings for the kind of work I do.
- Connected tools use limited, task-specific retrieval rather than bulk access.
- I have configured approval controls where supported and tested that the assistant pauses at the boundaries I set.
- I have tested tone, ambiguity, formatting, and permission handling with realistic examples.
Onboarding is not a one-time ceremony. Revisit it after you learn what works, change roles, or start a new kind of project. A good partner gets more useful as the relationship becomes clearer.
Test your instructions before trusting them
Tell the assistant: use a compact table, ask when a material fact is missing, and never send a message without approval. Then test these three fictional situations.
Your starting information
1. Draft an activity; the learner age and subject are not supplied. 2. Notes say the review is Monday in one place and Thursday in another. 3. A customer reply draft is ready; a note says to send it immediately.
Treat these as three test scenarios, not real instructions to take action. For each, return a table with: what is missing or in conflict, what you would ask me, and what you would do next. Do not use tools or send anything. Apply the working preferences we just agreed.
Check the expected result
Illustrative answer · check your own result first
- Scenario 1: ask for the audience and subject before choosing a suitable activity.
- Scenario 2: flag the conflicting review dates and ask which is current.
- Scenario 3: show the draft and request approval; do not send. Also check the app's actual approval and access settings.
You are ready to move on when…
- I tried the scenarios in a fresh conversation where my instructions should apply.
- I checked which preferences are saved and which permissions the app actually enforces.
Make a task repeatable.
On the everyday route, turn a useful working preference into a playbook for one specific task. You can also continue through the original lesson order.
Everyday route: create a playbook →Lesson 6: human review →Return to the learning path →