An Assistant That Knows You — in 20 Minutes
An Assistant That Knows You — in 20 Minutes — Your Personal AI Assistant: Build One That Knows Your Context, No Code Needed. Why the answers feel average; Your one-page About-Me note; Turning the page into 5–7 instructions; Fictional worked example: Ruth’s seven lines; The before-and-after test; What comes next, and the limit of a first setup. Practical…
Demo edition. The materials are undergoing editorial review.
Topics covered
- Why the answers feel average
- Your one-page About-Me note
- Turning the page into 5–7 instructions
- Fictional worked example: Ruth’s seven lines
- The before-and-after test
- What comes next, and the limit of a first setup
Why the answers feel average
Ask an assistant, “How should I plan a busy week?” Without context, it may suggest priorities, calendar blocks, and breaks. The advice may be sound, yet it cannot account for night shifts, caring duties, a tight budget, or your preference for short answers. The assistant can work only with the context available to it, so repeating your circumstances can improve an answer but costs time.
The goal of this lesson is to spend about 20 minutes creating a one-page About-Me note, reducing it to 5–7 standing instructions, and testing them on one real question. This does not promise perfect personalization. Instructions establish useful defaults, but facts still require checking and changing circumstances still require clarification.
Your one-page About-Me note
An About-Me note is an editable reference for you, not a full biography. Record six groups of information: who you are and your schedule; who you live with or care for; fixed limits involving time, budget, or food; 2–3 priorities; the preferred length, format, and language of answers; and what the assistant should avoid doing.
Keep only details capable of changing a useful answer. Do not include passwords, document numbers, exact addresses, other people’s secrets, or another person’s medical details. For planning, “I accompany my older father to appointments” can be sufficient; his diagnosis and documents are unnecessary. Check the chosen service’s data settings before storing personal context there.
Turning the page into 5–7 instructions
Select 2–3 stable circumstances that regularly affect decisions, then use the remaining lines to specify how the assistant should answer. Each instruction should describe one observable behavior: “Give the conclusion in 1–2 sentences first,” “Then use steps or a table,” “Ask up to 3 clarifying questions when information is missing,” or “Say directly when a plan does not fit the available time.”
As of November 17, 2026, ChatGPT provides Custom Instructions under personalization settings; Claude provides account-wide Profile Preferences and separate Project Instructions; and Gemini provides Instructions for Gemini for eligible personal accounts. Names, availability, and behavior change, so consult the current help page for your service. Notably, current ChatGPT documentation says updates apply immediately across all chats, so the older advice to always start a new chat is no longer universal.
Fictional worked example: Ruth’s seven lines
Ruth is an entirely fictional teaching character. She is 45, works as a nurse on 12-hour shifts with two days on and two off, and has two teenagers aged 14 and 17; the older one is applying to college. She also accompanies her 76-year-old father to appointments and studies online during evenings off shift.
Her instructions are: 1) plan around shift days; 2) account for two teenagers’ commitments and her father’s appointments; 3) allow about 2 evening hours for study on days off; 4) begin with 1–2 sentences, then use steps or a table, staying under 200 words unless asked; 5) ask up to 3 questions when needed; 6) identify an unrealistic plan directly; 7) omit motivational talk. Children’s names, addresses, and medication lists are absent because the stated task does not require them.
The before-and-after test
Choose one genuine question from the coming week. First, disable or temporarily remove the instructions and save the answer. Then enable your 5–7 lines and repeat exactly the same request. If the service does not promise to refresh the current conversation immediately, use a new chat. Compare four observable features rather than relying on a general impression: whether your constraints were used, whether the format fits, whether the answer invents assumptions, and whether the proposed plan is feasible.
In a fictional test, Ruth asks, “How should I plan next week? I have three shifts.” Before customization, she receives general planning tips. Afterward, she gets a suggestion to reserve shift days for sleep and dinner, move errands to two days off, and check the older child’s application deadline. This does not prove the second answer correct: Ruth must still provide the shift dates and check her calendar.
What comes next, and the limit of a first setup
Seven lines are only the initial configuration. A fuller assistant needs controlled memory, tested instructions, separate projects for different parts of life, short reference files, named routines, and rules governing access to external data. It also needs checks for excessive agreement, boundaries around major decisions, regular upkeep, and a portable “Assistant Passport.”
These principles can transfer between services, but their exact features, plans, and menus are not identical. A standing instruction is a default, not a guarantee: a current request or higher-priority service rule can change the result. Common first-setup mistakes include storing unnecessary personal data, writing vague wishes such as “give better answers,” combining conflicting rules, and judging the new setup with a different test question.
Practical assignment
Your first 5–7 lines
Continue to learningCourse outline
- An Assistant That Knows You — in 20 Minutes
- Five Layers: Where Everything Lives
- Memory Under Control
- Instructions That Actually Get Followed
- Projects: One Area of Life, One Workspace
- A Reference File Your Assistant Can Rely On
- Specialized Assistants: One Role, One Job
- Routines: One Phrase Instead of an Explanation
- An Assistant on a Schedule: Regular Check-Ins
- Voice and Phone: Capture on the Go
- Email, Calendar and Files: Access on a Leash
- Sycophancy: When Your Assistant Agrees Too Easily
- Big Decisions: Your Assistant Prepares, You Decide
- Keep Your Own Head: What Not to Hand Over
- When Your Assistant Gets It Wrong: Fix the Layer
- Upkeep and Moving: Your Assistant Passport