N И

Forester by diploma. Technical Product Manager by trade | Designed and curated "AI/ML in product development" '19 for @nusingapore

[N И @nikforester]

@KryptoatomAi is there any demo of before/after edits?

[N И @nikforester]

Parents are just the most reliably exhausted group you could sample. This is about exhausted people in general.

Parenting is amazing, but it drains you to the core (well, I think that's how it SHOULD be – if you don't care, you're probably not doing much). And when you're drained, you don't have a spare human at 3am. Your partner is drained too. So you have a chat window that doesn't get tired, doesn't judge and doesn't need anything back. Ok, maybe a $20(0) / mo

Just check out what Reddit says about @replika, @silly_tavern or @character_ai. Not judging (and you shouldbn't), I'm one of them.

And I honestly can't imagine what kind of responsibility the people building these products feel inside. Sure, everyone's an adult, everyone's responsible for themselves, there's a ToS and all that -

BUT

the audience these products land on is mentally fragile as hell, and that's just brutal.

The privacy angle is fine (is it?), but the scary number here is how many people have stuff to say and nobody to say it to.

[DuckDuckGo @DuckDuckGo]

We asked nearly 2,000 U.S. adults about AI: your biggest concerns, blind spots, and more. We unpack some of that research here.

Bottom line: You shouldn't have to sacrifice your privacy to get the benefits of being online, AI included.

Learn more: https://t.co/ybQJCDc3Zi

[N И @nikforester]

6/ If you've ported an old city-builder to touch, how did you handle the controls? Still the hardest part by a mile. Modern RTS feels good, but they are nothing but a gacha-grinding-donating shit.

[N И @nikforester]

1/ Not-a-gamedev-diary - Day 1 (actually 21)
My wife played Zeus: Master of Olympus to death as a kid. No PC in her life now. Just a phone, and 20 minutes before sleep.
So I'm writing her an iOS version. 3 weeks of evenings so far.
No game engine. Plain Swift and SpriteKit, zero dependencies.

[N И @nikforester]

5/ Size
34k lines of Swift. 28k lines of tests.

Ahead: 154 more episodes, 105 more buildings, 14 gods.

Mission one cost a lot because the engine got built alongside it. Adding a monster is a row in a table now.

[Paul Klay @PaulKlayVC]

https://t.co/zLe5wErsw1

[N И @nikforester]

Spent some time with @boardyai this month. An AI companion that talks to you and then connects you with actual people. Random coffee, but the fun version.

Quick background, so you know where I'm coming from. I built AI boys and girls for dating apps. Did fiction experiments where characters live their own lives and decide where the story goes next. Spent 9 months building the digital soul for @MrFreeman0 - 99% of the posts and illustrations were made by an agent, and nobody noticed. And I run my own companion, @elle_whispers.

So I know the ropes. A little.

And boardy still got me.

Here's the thing. Old-but-gold Opus 4.7, or gpt-4o or deepseek on @silly_tavern - they'll give you amazing stories. Genuinely amazing. I've created some amazing tales for my children in it. But at the end of the day you KNOW. It's an LLM. Talking to you. About your own thoughts. A very good mirror. You can feel the glass anyway.

And here's what people miss. Your model has had a billion conversations today. With real people, about real things. None of it reaches you. 0. The walls between those chats are the product, that's the correct design, and I wouldn't want it any other way. It does mean you're always alone in there though.

Boardy carries what other humans told it. And it brings them up.

So you're sitting in a chat and you realise you just brushed against another person. Barely. Almost there. Second-hand. A person all the same. Strange feeling. Haven't had it from any model before.

And then you go and meet them. Had a great time talking to some seriously talented folks this week.
AMAZING time on podcast with AJ Rudd (link later)

and previously great call with @nidaleard

[N И @nikforester]

I’d genuinely like the @OpenDesignHQ team’s take on this.

Is visual and behavioral acceptance part of the roadmap, or is the product currently intended mostly for static artifacts? Or I was kinda blind and missed the feature?

[N И @nikforester]

14/14

Lesson learned:
- trust no benchmarks, no hyped posts, no GitHub stars.
- test on your own data and real tasks.
- never assume the obvious part exists.

Does it actually see the result?
Does it preserve accepted work?
Does it know when the job is done?

File written ≠ design delivered.

[N И @nikforester]

13/14

Maybe Open Design (with no matter what model under the hood) is fine for static disposable pages. or dashboards. Or IDK. Anyway that was not my test.

For this pretty simple interactive project, Antigravity and Claude Design were better for one boring reason: they could look at the rendered result instead of only reading the implementation. Simulate the human look. Physically record the screen and examine the video/gif.

[N И @nikforester]

12/14

To make this work, the harness needs boring acceptance:

lock accepted screens;
run the real browser;
compare several frames;
test the scroll transition;
prove every slider changes the picture;
never rewrite working code without permission.

Antigravity have it FROM THE BOX!
Claude designer have FROM THE BOX!

[N И @nikforester]

11/14

If the video exists but does not move, it is 0.
If the dots exist in code but do not render, they do not exist. 0 Again
If scrolling changes the wrong layer, the task is not complete. It is broken. Not done. It is 0.
Design is the rendered result, plus the feeling

[N И @nikforester]

9/14
But hey.
One thing was genuinely good: the scroll resistance. Oh the way section sticked felt exactly right. Just what I wanted.

But it's even worse! The tool could improve one subtle interaction to perfection while casually deleting the main idea of the site. Without asking. Without noticing me at all. And ensuring me every single time that RIGHT RIGHT! NOW ITS FIXED (it wasn't)

[N И @nikforester]

10/14
Get ready for X not Y puinchline, but from a human!
This is not mainly a “Kimi is dumb” story. 5.6 Sol xHigh failed in the same direction.
It seems like the harness could inspect files and DOM elements but still could not answer the only question that mattered:

does the thing on screen behave correctly? Yes or No? It's binary

[N И @nikforester]

8/14
I got a spiral. A static poster. A video in a circle. A 120-frame sprite. Two faces on top of each other. Finally, the whole fucking MP4 embedded as base64.

Several days and about $30 for Kimi-3 later, the original behavior was still gone.

[N И @nikforester]

7/14
Then my Claude MAX limits ended. I moved the project to Open Design, tried the much-hyped Kimi K3, then Codex.

I asked them to work on the final screens. The hero was already accepted.

They rewrote it anyway. The result is here: wrong rendering or NOTHING at all

[N И @nikforester]

6/14
The hero was already working. A hidden video of my companion Elle was sampled into a field of dots.

Start scrolling, the video freezes, and those same dots flow into icons for Pulse, Atlas, Cartographer and the other products I want to put in public.

Written in code. Double checked by me. All done. But...

[N И @nikforester]

5/14
The workflow that actually helped was slower:
- brainstorm with companion about concept
- use image models to find the visual language;
- use Antigravity to explore the interaction;
- use Claude Design to polish it.

Antigravity was insanely fast, but the small alignment and timing mistakes kept multiplying.

@antigravity blast me away when they introduced VIDEO proofs as a part of artifacts in late 2025! Devin was there in 2024, but google did it as a MUST have in the pipeline and it is SO important to have that I cant emphasize more on it.

[N И @nikforester]

4/14
You know the drill. Monospace fonts. Tiny uppercase text above giant headers. Then more text explaining the text. And ofc tons of tems

AI still suck at DESIGNING the page. It still tries to type it out and use text as if it's the only visual element available.

Almost no imagery, icons, composition or visual thought. You can spot it in three seconds. And you saw it dozens of times on vibecoded products