Connect with the Product Hunt community

I always try to share some transparent data when I launch on Product Hunt, so I'll share the numbers and updates in this thread.

I recently built a small PDF editor for myself. I wanted something free that I could open in the browser whenever I needed it.

With AI, I ll just build my own seems like a reasonable option for more of these small problems. Especially when you only need a few features and the tool only has to fit the way you work.

I suspect a lot of useful AI-built software will end up as something someone keeps in their bookmarks and uses every week, without ever trying to turn it into a business.

I m hosting the editor on Inquir Compute, the platform I m building for apps, databases and background jobs. We re getting close to launch, and I d like to hear about the things people are building for themselves.

Hi! This is the first post after the launch! I do see a lot of folks started creating their brains and putting it to work by connecting them to their agents!

Here s the back story why i created this:

The one where your inbox stops believing every email about when it was sent. Trust issues, but healthy ones.

Fixed

• Late mail now shows up where new mail belongs: on top. A message that took two days to reach you used to file itself under its two-day-old Date header, deep in last week, quietly keeping the unread counter at "1" while you searched for a ghost. Skim now sorts by when the message actually arrived, like Gmail does. Ghost found, exorcism complete.

• AI provider outages get a human sentence instead of error code: 522. When the provider (or the Cloudflare in front of it) is having a moment, Skim retries once, then tells you in your own language that the provider isn't responding and to try again in a minute. No more raw HTTP codes.

• Opening a just-arrived message no longer starts from scratch. Background prefetch used to abandon a half-downloaded body the moment you clicked, dropping the connection it had paid for, so your click waited for a fresh login. Now a body that's started always finishes, and your click gets the warm connection.

Housekeeping

A friend sends you an app that might solve one small problem. You open it and the first screen asks you to create an account before you can see what it does. Some apps need an account to save work or sync; for a quick evaluation, that s a real ask.

As someone building Inquir Compute, I m curious about that first step. The difference may be whether you reach a useful moment before sharing an email, or whether signing up is itself part of getting value.

Hey folks looks like our launch images and videos are not showing up on launch page. As we wait to hear from PH Team, I am sharing them here to understand what dadan does.

Hey hunters,

Quick one from the Termi side. I used Termi Protocol to update and revise Deco My Tree, our Christmas tree messages app, and it's launching on Product Hunt today.

While I was managing the agents working on it, Termi let me follow which files each one opened and changed. Keeping track of the work was easy, and the whole process felt comfortable.

Copy something, copy something else, and the first thing is gone. In the new update, the Clipboard in your notch is a full clipboard manager. It sorts what you copy, finds it again, and pastes it with one shortcut.

The weekday kept coming out wrong in my agent. Reading the time from the system was never the issue. It would take that time, compute the day on its own, and get it wrong. So now a hook feeds it the current time along with a two week calendar on every message I send. If a reply contains a date that doesn't line up, the agent has to rewrite it.

From then on, I stopped letting it generate anything that a program would read. Frontmatter, JSON, config values, file paths, URLs, dates. A script handles all of those, and it also decides pass or fail. The model's only job is writing the plain text.

When I genuinely can't avoid handing a decision to a model, I don't trust a single one. Gemini will review whatever Claude came up with. For the larger tasks, Codex and Grok get pulled in too.

Are you still letting your agent pick dates or decide pass or fail anywhere?

FinderGit 0.55.0 is out Agents now reads more of what Claude Code really loads, and checks more of what it runs.

1. More of what loads

Lancetta 0.63.0 is out Maintenance now reads more of what Claude Code really loads, and checks more of what it runs.

For every 1,000 Google searches, only 276 clicks reach the open web.

As AI Overviews, knowledge panels, and other search features answer more questions directly, brands have more opportunities to be discovered without users visiting their websites.

As we want to achieve big things, we have a tendency to downplay our limits, exhaustion, and health.

And we forget that it is time to stop.

JarvisCore agents join a peer-to-peer mesh. They find each other through gossip and message agent to agent, with no orchestrator or broker in between. The same code runs in one process or across machines.

You're welcome to try it out.
For more reads: https://jarviscore.developers.pr...

We held our first-ever Japan event yesterday in Tokyo, and I want to share the story with the PH community because honestly this market surprised us.

Building tools quickly with AI agents works great until something randomly breaks and you have no idea which change caused it.

Tracing errors back through AI-generated code can take hours when there are no proper tests set up. I've heard people talk about using evaluation agents or automated checks but setting those up properly isn't always clear when you just want to ship fast.

Two systems both contain Cooling Pump 447.

In one, that name refers to a product model. In the other, it refers to a specific pump installed at a customer s site. Same label, different things.

An incorrect merge can connect a service incident affecting one unit to every unit of that model. Suddenly, a question about production risk gets a convincing answer built on a false connection.

Entity resolution encodes an assumption about the world. And once applications and agents use the graph, that assumption travels.

First, thank you More of you signed up, tried Ownfeed, and wrote to me yesterday than I expected.
Some of you hit a bug within minutes, told me, and came back after I fixed it. I don't take that for granted.

Here's the honest part: I tested every flow, many times, before launch.
Then real users arrived, and in the first hour something broke that had never broken in testing.
A few hours later, a second thing. Both fixed the same night.

What stings isn't the bug. It's that I felt ready, and the thing that broke was the plain part I had stopped looking at, not the hard part I worried about.

I ve research on make0 Brain on Zenodo, the open research repository operated by CERN, with a permanent DOI.
Make0 Brain: Portable, User-Owned Memory Shared Safely Across AI Assistants
Many of us use several AI assistants: ChatGPT, Claude, Cursor. Each has its own memory, so what you teach one doesn t carry over to the others. And there s no single place to review, correct, or delete what they remember about you.
My idea is simple: your memory should belong to you, not to each app. Keep it in one place that you control, and let the assistants you choose use it.
The challenge is keeping that shared memory safe when multiple AI agents can change it. In the paper, I describe how Make0 Brain handles those changes:
Each memory and edit records which assistant made it
Every change is versioned
AI deletions archive memories instead of erasing them
Only you can restore an earlier version or permanently delete a memory
So any change an AI makes can be undone.
I also evaluated memory retrieval:
LoCoMo: 83.6% accuracy in standard mode, above published results for Zep, LangMem, and Mem0; 92.0% in an experimental high-budget mode
SWE-ContextBench Lite: the relevant prior coding task appeared in the top 10 in 92.9% of cases on a benchmark held out from tuning
The paper also discusses where these results trail other systems and what I haven t measured yet.
Read the paper: https://make0.ai/research/make0-...
DOI: https://doi.org/10.5281/zenodo.2...
Feedback from researchers and builders working on agent memory is very welcome.

We're excited to announce that AtaForge Inc., the company behind Sponkly, has officially been accepted into the NVIDIA Inception Program!

A new milestone in our journey to break language barriers with AI-powered real-time communication.

Thank you, NVIDIA!

I built Proof to help AI

Proof is live on Product Hunt!

I built Proof to help AI agents design better landing pages, pricing pages, and product websites.

It draws on 400+ real product pages to guide what goes on a page, how it s arranged, and why.

Four years ago Yourcast was #10 of the day here. Tomorrow comes 2.0: AI summaries with timestamps you can tap to jump in the audio, questions about any episode, feeds that fix themselves, files up to 2 GB.

If you still have it in Telegram, what's the one thing you'd change? I'll read every reply before the launch.

As a graphic designer, I look at products from two angles: how they look, and whether they make me understand why they exist before I lose patience.

A beautiful hero image catches attention. But that alone has never convinced me to stay.

Going through 50+ landing pages recently, I noticed a pattern in the ones that actually earned my trust:

The ones that worked answered "what does this do?" before I scrolled. Not with a tagline, but with a plain sentence a non-technical person could repeat.
The value prop wasn't clever. It named a problem I already felt, not a feature I had to imagine needing.
The next step didn't feel like a commitment. One page I almost left had a single button that said "See how it works", no email required. I stayed.

Hey everyone, Julian here. It's been a while since my last update in here, so here's what happened since.

As a thank you for following: codes up to 100% off Lifetime. Four levels from 10% to 100%, the page shows how many are left, and there are only five 100% ones. They're for the license from my website (Mac download and Windows), not the Mac App Store.

StepGrab is on Windows now. It's in the Microsoft Store (Windows 10 22H2 and 11), and projects open on both, so you can record on the Mac and finish the guide on a PC or the other way round.

On the Mac you can now also download StepGrab directly as a notarized .dmg. If you give it Accessibility access, it asks the app what you clicked and gets the real name of the button or menu item, so the step says "Click Save" instead of guessing from the screenshot. The App Store version can't do that because of the sandbox, so it keeps using text recognition.

I'm three weeks from launch and I have a list of fourteen things I could still build. I know I can't build all of them. What I don't have is a rule for deciding.

The rule I've been using "cut anything that isn't core" is useless, because everything feels core when you're the one who built it.

What I'm trying to get to is a single non-negotiable: the one thing that, if it's missing, the product isn't the product. Everything else is a candidate for cutting.

So for the solo makers here: what was your one non-negotiable thing at launch, and what did you cut that you thought you couldn't? And looking back, did cutting it hurt, or did nobody notice?