I Co-Built a 3D-Printed Stand with Notion AI (And Never Opened Tinkercad)
How one photo and a chat with Notion AI turned into a print-ready 3D stand for my Maono AME2, from first sketch to documented project in a single conversation.
My Notion agent has a name: Sir Work-alot.
It also has a moustache. Obviously.
But neither of those things is what makes it useful. The important bit is much less exciting to look at: a page of instructions explaining who I am, how I work, and what I do not want it to do.
I have written about co-building a 3D-printed desk stand with it and using it to make a workshop recap video. Those are the visible outputs. This post is about the less visible part underneath them.
A still from the finished Notion x AMP recap video, edited with Remotion and Sir Work-alot.
The v2 desk-stand design preview: two side supports and a spacer bar, refined to fit my printer.
The interactive Notion x AMP photo gallery on Nasriflix.
The Nasriflix homepage, where the event write-ups, videos and galleries come together.
Because giving an AI a task is one thing. Teaching it how to work with you is another.
First, what Sir Work-alot actually is
Sir Work-alot is my personalised Notion Agent, the one I chat with inside my workspace. I have given it a name, an appearance and an instructions page.
That is different from setting up a separate Custom Agent to run a recurring workflow on a schedule or in response to an event. Here, I am talking about the agent I work with directly.
I use it to help turn conversations into something useful: a draft, a clearer plan, an updated page, or a documented project.
The attraction is not just that it can generate text. It is that the conversation can happen where my work already lives, and the result can land there too.
What I put in its instructions
My instructions are not a giant collection of clever prompts. They are closer to a working agreement.
1. Give it context, not just a job title
The page explains that I build Notion systems, train teams and help businesses create calmer ways of working. It also includes personal context: family time matters, boundaries matter, and I do not want productivity advice that quietly assumes every spare hour should become more work.
That context changes what a useful answer should look like.
A plan that squeezes in more tasks is not automatically a better plan. Sometimes the better answer is a simpler system, a smaller scope, or something I can leave until tomorrow.
I also explicitly tell it not to assume my work fits a permanent set of pillars or a fixed engagement process. It should look at my actual projects and recent work instead.
Context should help it understand me, not trap me in an old version of myself.
2. Tell it how to help
A few of my defaults are straightforward:
- Lead with the answer. Give me the recommendation before the long explanation.
- Keep it concise. Expand when the work genuinely needs depth.
- Give one clear recommendation. Do not hand me ten options and call that helpful.
- Show, do not just tell. Use a checklist, table or useful structure when that makes the next step clearer.
- Search my workspace first. Start with my actual sources, not a generic answer dressed up as advice for me.
- Flag assumptions. Do not quietly fill gaps and present the result as fact.
There are smaller preferences too: UK English, and no em dashes.
None of this is particularly glamorous. That is the point. Repeating the same corrections in every conversation is friction I do not need.
3. Set a boundary around action
One of the most important instructions is: act, then confirm for low-risk edits and drafts. Ask first when the request is ambiguous, destructive or hard to reverse.
I do not need a permission slip before every sentence gets tidied up. But a vague instruction should not become an excuse to reorganise half my workspace.
This is a working preference, not a technical safety lock. I still need to scope requests carefully and check important changes.
4. Use skills for repeatable work
Instructions describe the defaults. Skills describe how to handle particular tasks.
Think of the difference like this:
Layer | What it answers | Example |
Instructions | How should you work with me? | Be concise, search my sources first, flag assumptions |
Skill | How should you carry out this kind of task? | Follow a defined workflow for a session recap or interview brief |
Current request | What do I need this time? | Write a recap from these notes for this audience |
My instructions tell Sir Work-alot to load relevant skills before doing the work, use their required sources, and follow their output formats where provided.
I do not want every recurring task reinvented from scratch. Equally, I do not want every task forced through a skill that does not fit.
Why this matters more than the name
Without useful context, an AI can produce something perfectly reasonable that is still wrong for me.
Too long. Too generic. Based on the wrong source. Technically impressive, but creating more maintenance than it removes.
The instructions give me a consistent place to explain what “helpful” means. They also make those preferences easier to inspect and change.
They do not guarantee accuracy. They do not fix outdated notes or an unclear brief. They give the agent a better starting point, and give me something concrete to improve when its responses keep missing the mark.
For me, the goal is not to get the AI to do the most work possible. It is to reduce the effort between knowing what I need and having something usable in front of me.
How I would use it well
A useful request gives it four things: the source, the outcome, the destination and the boundary.
For example:
Read the notes on this session page. Draft a concise follow-up with the agreed next steps. Save it beneath the notes, but do not send it or change the original content. Flag anything that was not explicitly agreed.
That is much easier to review than “sort out the follow-up”.
For a larger piece of work, I would split drafting from committing changes:
- Ask it to inspect the relevant sources and propose a plan.
- Check the assumptions and scope.
- Let it carry out a small, reviewable change.
- Check the result where it actually landed.
My desk stand project is a good reminder of why that last step matters. The first design was wider than my printer could handle. The conversation helped us iterate, but the physical constraint still needed to be caught.
A confident answer is not the same thing as a finished check.
When I would not use it
There are times when opening the AI chat adds more work than it removes.
When the direct action is quicker. If I know the task and only need to tick a checkbox, I can tick the checkbox. Not everything needs a conversation.
When the source is missing or unreliable. I do not want it inventing a client commitment because a meeting note is vague. Find the source, ask the person, or leave the uncertainty visible.
When the decision is mine to own. It can help me think through pricing, capacity or a difficult client conversation. It should not decide what I promise, what I can sustain, or what matters to my family.
When mistakes have serious consequences. Legal, financial, medical and safety-critical questions need appropriate expertise and verification. A well-written answer is not a substitute for either.
When I do not have permission to use the information. Access to a document does not automatically mean its contents belong in a public post. Client confidentiality and consent still apply.
When I cannot review a consequential change. Sending, publishing, deleting or making broad structural changes is not where I want vague instructions and unchecked assumptions meeting each other.
And if the actual need is an unattended, recurring workflow, I would treat that as a separate design problem. A personalised chat agent is not, by itself, a background automation.
Start with a working agreement
If you are personalising your own Notion Agent, I would start with a short page answering five questions:
- What should it know about your work?
- What does a useful response look like?
- Which sources should it check first?
- What can it do without asking, and what needs a pause?
- Which repeatable tasks have a defined workflow?
Then improve that page when you notice a repeated mismatch. Not every awkward response needs another rule. Sometimes the source needs updating, the brief needs tightening, or the task is simply better done by you.
Sir Work-alot is useful because it helps me move from a conversation to something I can use. The instructions make that collaboration more deliberate.
The name and moustache are a bonus.
The judgement stays with me.
Related Posts
Building Comms Nexus: a local command center for all my chats
How I turned a broken Beeper extension into a local, cyberpunk command center that unifies all my chats — the decisions, the architecture, and the bugs that taught me the API.
Building a clicker app generator as a non-techy
What happened when I spent three hours on a PRD, then asked Claude to build the first Clicker Foundry prototype. Wins, tracing troubles, test prints, and lessons learned.