I Built 45 AI Tools in Less Than a Year. They All Fall into 8 Patterns.
I used to teach at the college level (Go Coogs!) and miss those days. The energy of a campus along with the infinite possibilities was intoxicating. New students stepped onto campus with ambitious ideas and their adult lives in front of them.
Over the next few years, they start to settle in and find their path. This is most obvious in their major, but also in their friend groups, clubs, sports, and extra curricular activities. That large world that they started in on day one begins to get smaller and more defined. The energy and excitement of the campus is still there but somewhat muted as life becomes more predictable and structured.
What does this have to do with AI? Everything.
Those of us who work in AI, not those who dabble, felt the same excitement when they began building agents, skills, tools, and other applications. The excitement of working side-by-side with a program in near-human conversation was exhilarating.
Then time passes and the thrill changes to systems, structures, knowledge, seeing things others don't, and realizing most of your work falls into patterns.
Recently, I was asked to audit every AI tool I'd built, or had a heavy hand in since joining Zinus less than a year ago. The honest first reaction was to count each one. The number came back at 45. My second reaction was the one I want to write about, because it's the useful one: when I actually looked at the list, those 45 tools weren't 45 ideas. They were 8 patterns built similarly but for different purposes.
The number is the least interesting thing on the list
Forty-five sounds like a lot. It looks good for a LinkedIn post. But the count tells you nothing you can use. It doesn't tell you what kind of work AI is actually good at, and it doesn't help you build your first one. Also, real practically, anyone can "build" something with AI in minutes so that large number doesn't say much.
What helped me was grouping them. Once I sorted the 45 by what they fundamentally do rather than which team they belong to or which platform they live on, the same shapes kept repeating. A tool I built for marketing turned out to be structurally identical to one I built for customer support. Different inputs, different audience, exact same mechanics underneath.
If you're staring at AI wondering where to start, you don't need 45 clever ideas. You need to recognize the rhythms you already live in during your week.
The 8 patterns
1. Turn a mess into a clean format. You have something unstructured, for example, a Word doc, a raw export, plain text, and you need it in a strict format every time. HTML. JSON. A knowledge-base article. A good chunk of my tools do only this. My Word-to-HTML converter and alt-text generator are both public examples. The work is boring and high-volume, which is exactly why AI is good at it and exactly why nobody wants to do it manually (I can't believe I used to manually write HTML - sometimes I miss the simplicity of it, but now, that time of my life has passed).
2. Run the report that runs every month. Any recurring deliverable on a fixed cadence: a monthly deck, a quarterly summary, an executive one-pager. You do the real version once by hand so you understand it, then you encode the format so each cycle becomes assembly instead of construction. This is where my 10-80-10 Model earns its keep: you own the inputs and the final read, AI does the middle.
3. Get a fast answer you can defend. A decision is on the table and someone needs analysis before the meeting. The page-grading tool I open-sourced is one of these. So is the SEO research that used to eat 16 hours and now takes 15 minutes. The word that matters here is defensible as the output has to show its work, because you're going to have to stand behind the number. We could spend all day here as the amount of AI generate confidently incorrect files is way too high.
4. Keep everything on-brand without thinking about it. Color, typography, voice, logo rules. Encode the brand once and every artifact comes back consistent without a manual style pass. This is an area where Claude Skills are exceptional and you can share them across your organization with a couple of clicks.
5. Let AI take the first pass at support. This is where I spend most of my day job, and it's the highest-stakes pattern on the list. When people picture AI in customer support, they picture the part the customer sees: the chat window, the voice on the line. That part gets all the attention. The part that actually decides whether any of it works is the knowledge sitting underneath it. An agent is only as good as what you've taught it, and a thin or contradictory knowledge base doesn't make the agent cautious, it makes it confidently wrong, which is worse. So I learned to flip the order. Build the knowledge layer first, get it clean and consistent, and treat the agent itself as the easy part that comes last.
6. Keep a big project from falling apart. This one took me the longest to respect, because it doesn't produce anything. Here is where I track my largest team projects, e.g., vendor negotiation, full year strategy, or deploying a new platform. There's no true finished deliverable at the end, no deck or report you can hold up. It's just a place that holds the goals, decisions, and open items for a messy, sprawling project so nothing slips and the details aren't trapped in my own head. I spent a while trying to pin a clean metric on it and finally gave up, because there isn't an honest one. Its value is invisible by design. It shows up as the deadline I didn't miss, and the decision I could actually explain three months later when someone asked why we made it.
7. Write the email faster. It's cliche to use AI to write emails, but this is helpful. And here is where it helps the most: with international teams. I built a Claude Skill that recognizes recipients and their native language. Then the email is converted to be more friendly to translation software since I know they will use it, i.e., remove jargon, American idioms, break up long paragraphs, etc. Essentially, I write the first draft and the Skill coverts it to be ingested by translation software.
8. Check the AI's work. Two of my most-used tools do nothing but audit other tools' output by flagging fabricated specifics and numbers that don't match the source of truth before they reach a leadership report. The more AI you run, the more you need this. AI that sounds right and is quietly wrong is the most expensive kind. Of all the tools I use the most, my 'verification-output' is the one that is used daily.
About a third of the list is dead
Here's the part that doesn't make the highlight reel. When I finished the audit, roughly a third of those 45 were retired, superseded, or flagged for removal.
One of them stung a little. I'd built a tool to write detailed, structured prompts for AI image generation. It worked, and the reality is I had spent over a year on it since image generation still wasn't great, but then it finally was solid. It produced genuinely good marketing images. Six months later, the image models improved to the point where the elaborate prompts barely mattered anymore as you could describe what you wanted plainly and get the same result. The tool wasn't wrong. It was temporary.
That's the lesson hiding inside the retirement list. The tools are disposable. The patterns are not. I used to tell my students, "never get attached to anything you create, except your children." This was usually met with laughter, but it's true then and today, your professional creations should be burned to the ground every once in a while, especially, if they don't add value anymore. With AI, a specific build will get obsoleted by a better model, a platform change, or a native feature that didn't exist when you started. But "turn a mess into a clean format" has been useful since the first one I built and will be useful after this one dies too.
Count your patterns, not your tools. The tools expire. The patterns compound.
If I'd fallen in love with the image-prompt tool, its obsolescence would feel like a loss. Because I think in patterns, it was just one instance of pattern #1 retiring while the pattern itself kept working everywhere else.
Start with the pattern, not the tool
If you're trying to get real value out of AI and you don't know where to begin, skip the brainstorm. Don't try to invent something clever.
Look at your actual week and find the pattern:
Where do you reformat the same kind of mess by hand? (Pattern 1)
What report do you rebuild every cycle (monthly, quarterly, etc.) from scratch? (Pattern 2)
What decision keeps stalling because nobody has time to do the analysis? (Pattern 3)
Pick one. Just one. Do it manually first so you actually understand the steps because AI scales a process, it doesn't invent one for you. Then document it, and be boring about it, "step 1: I opened X program, step 2: I clicked on Y button, etc." That's a single tool. Now you've matched one pattern to one piece of your real work, which is worth more than a list of 45.
I didn't set out to build 45 anything. I set out to solve one annoying problem at a time, and the patterns revealed themselves in the audit. Yours will too. You just have to stop counting the tools long enough to see the shapes.
You don't need more ideas. You need to recognize your daily patterns.
