Master WorkBuddy in Three Steps: Context, Memory & Skills for a Personalized AI Assistant

Step 1: Context — The Sticky Notes on Your Desk

Context is everything the AI can see in the current conversation: your typed messages, uploaded files, and pasted screenshots. It acts like sticky notes on a desk — cleared when you close the chat. The quality of context directly determines output quality.

The author contrasts two approaches: one user vaguely asks "help me process that sales spreadsheet, organize the important stuff" and gets nonsense; another drags the spreadsheet in and adds "summarize by month, calculate month-over-month for sales column, output to new file" and succeeds immediately. The difference: the first makes the AI guess; the second lets it see.

In WorkBuddy, "let it see" follows three patterns: @ reference files, drag screenshots directly, provide a template . A template beats a thousand words of description — feeding last year's formal report works better than saying "format formally" ten times.

Counter-intuitively, more context is not always better. The author found that after 30+ turns in one conversation, the AI became confused because it had to search through old "sticky notes" and picked the wrong ones. Solution: start a new session per task, like changing water for shrimp — stale water kills them.

Step 2: Memory — The Employee's File Cabinet

Context solves one conversation; memory persists across conversations. WorkBuddy uses a MEMORY.md file — a set of rules the AI reads before every task.

The author shares their actual MEMORY.md content:

Write WeChat articles: conversational, paragraph-based, zero AI tone; use dark-green theme; titles must provide multiple grouped versions; when advancing the pipeline I only send single-character commands.

The last rule is the most powerful. Previously the author had to micromanage every step; now a single character "continue" triggers the next step because the workflow is encoded in memory. One-character command works because everything was pre-written into the file.

Discipline: keep MEMORY.md under 3,000 characters. Overlong files cause the AI to miss key points. It stores rules, not life stories. Two levels exist: user-level (personal preferences, portable across projects) and project-level (project conventions, scoped to that project) — analogous to personal files vs department regulations.

Step 3: Skills — The Employee's Expertise

Skills are like apps on a phone. Without the docx skill, asking for a Word document yields messy headings and broken tables; with it, the same prompt produces a polished document.

Invocation is simple: prefix the command with @. Example:

❌ "Help me organize this proposal document" (free-form, random formatting)

✅ "@docx read this proposal, extract all headings and corresponding points, output as new document" (precise, one-shot)

Golden rule: always specify the skill in the instruction . Don't test the AI's intuition; explicit designation guarantees stability.

Beyond using built-in skills, you can create custom ones. The author built a WeChat writing skill packaging their style, material library, and banned-word list. Now a new chat with "write per old rules" invokes the entire personalized workflow.

That moment felt like not leaving programming, but moving 20 years of experience from code into memory files and skills.

Putting the Trio Together

Recurring weekly report format → write once in MEMORY.md, use forever. This month's data files → drag into context via @, cleared after the session. If the report process is fully standardized, wrap it into a Skill or automation so Monday morning it generates automatically; you only review.

Context = sticky notes (changed daily). Memory = file cabinet (accumulated slowly). Skills = craftsmanship (carried for life). With all three, the "intern who needs re-teaching every time" becomes an assistant who truly knows you.