Designing by elimination: how Improve with AI got made
Hatred is informative. It's also a terrible design brief. The story of how Improve with AI got designed, from a list of grievances to an actual feature.
I stare at the field and blinking cursor. It says "Type in a question...". I stare some more, my brain firing and thinking. I want to ask something, I know I want some kind of output but, what exactly?
Most AI tools open like a stage with no script.
It's almost like a person just standing there, waiting in the spotlight, and the audience is feeling awkward, wondering when the performance will start. Is this part of it?
I personally hate this. I find it genuinely paralysing. Tell me what this thing does. Tell me when I should use it. Tell me how. Give me a verb and a noun and a button.
AI is amazing. It can help us become more creative, it can make our lives easier. But it can also expose something in us that we don't like. We don't like not knowing what we want.
I am a grown adult with a job and a calendar and approximately four hundred unread Slacks. I do not have the emotional bandwidth to invent a use case from scratch every time I open a new tool.
Generating something from scratch can absolutely feel magical and that's the point. Drafting, translating, rewriting an entire article in a new tone, summarising a giant doc into something readable.
Those are jobs where the AI has clear permission to take a swing. It's something tangible. Do x so that y happens. You hand over the input, you get an output, you compare the two, you move on. The transaction is clean.
Reviewing with a sense of control
Reviewing an article you've already written is a different shape of job. You know you want it to be better, but you're not quite sure what, and you want control over it.
The article exists. You're emotionally attached to it, whether you like it or not. You might dislike or like what you've produced. You might feel that it needs a once-over, like every great author.
You've made a hundred small internal calls about voice, structure, where to put the callout, and yes, whether that joke is too much.
River Sloane
For this job, "magic" is exactly the wrong texture. There's no way you want it to simply improve your article and then move on. You don't want a wand. You want a checklist.
You want to see:
- what the AI noticed
- where it noticed it
- and why it would suggest it
all before any of it gets within ten feet of your draft.
And so when we started building Improve with AI, the very first thing on the list was "this should feel contained." Knowable. Boring in the good way. A tool with edges you can see.
Which sounds great, except "do not feel like a magic box" turned out to be a really bad design brief on its own.
We had a long list of things we did not want. We knew we didn't want a chatbot-shaped void going "amazing!! try anything!! so damn GOOD ✨" every time you pressed a button.
We knew we didn't want ungrounded, "your article could be improved" feedback that makes you feel gaslit. We knew we didn't want changes appearing without warning. We knew we didn't want a confidence-shaped trap that told us everything we wrote was brilliant.
Cool. Several things we hated. Zero things we'd actually built.
River Sloane
Trying to find something tangible
The hard part, the genuinely embarrassing part, was working out what the opposite of all that even looks like 🤔
"Not a yes-person" and "not magic" don't add up to a positive shape on their own. You can avoid every single thing on your hate list and still ship something joyless. Negative space is not a feature.
So we spent an unflattering amount of time trying to describe what we did want, in actual words, like adults. We wrote a lot of bad first attempts.
"Helpful but firm", "Honest but kind", "Like a really good editor". All technically true. All bland enough to serve at a continental breakfast.
The breakthrough was realising we didn't need a tagline, we needed a behaviour.
A specific picture of the thing happening. We kept coming back to the same mental image: a colleague who reads your draft, makes a small huff, points at the third paragraph, tells you what's wrong with it, and lets you decide what to do about it.
The friend who fixes your eyeliner for you, not the friend who lets you walk into the meeting like a panda. Once we had the behaviour, the decisions started writing themselves.
What, where, which first, when, how. Every step visible. Every step yours.
- What. Improve with AI's job is helpful feedback. You get a score. You see the suggestions, you read them, you decide whether to apply them. The output is information.
- Where. Suggestions that are located. Points to a paragraph with a "Near:" label. There is a gigantic gap between "your article could be improved" (cheers, life-changing) and "your second paragraph assumes context the reader doesn't have yet" (something you can actually fix before lunch).
- Which. Suggestions come tagged Low, Medium, or High. Because a useful colleague tells you what to fix first. They don't hand you a list and wander off whistling.
- When. Nothing changes until you say so. Every suggestion shows you a preview before you apply it. Apply one, apply all (YOLO), apply none.
- How. It works in both directions. Browse suggestions in the sidebar and jump to the right spot, or click a highlighted block in the editor and the suggestion comes to you. Good coworkers meet you where you are.
This is what I mean by contained. Every suggestion has a category, a severity, a location, a preview, and an off-switch. Nothing in the feature happens that you didn't see coming. No magic void.
Design by feeling, then refine
I guess the thing I'd most want a designer or PM to take from this (as if I'm some kind of PM guru) is that having a strong negative reaction to existing tools is not, on its own, a design brief. But hatred is informative.
It's highly unlikely that I'm the only one that has strong feelings about blank boxes outside of the obvious all-in-one AI tools like Claude and ChatGPT.
You still have to do the slow, slightly embarrassing work of writing down what you actually want, deleting it because it sounds rubbish, and trying again until something rings true.
We did. Eventually.
It turned out to look a lot like a coworker who reads your draft and tells you the truth, in a specific place, at a specific moment, with a clear button to push. Boring in the good way. Contained on purpose.
Become an expert in all things Knowledge Base with our monthly newsletter. No spam, just expert content, delivered.