Connect with the Product Hunt community

TL;DR: every fact in our graph is its own node, with properties on it. The vocabulary behind those nodes is RDF/OWL. Neither camp would call this clean, and Wikidata is not strictly one or the other either.

Here is why we ended up there.

What forced it

Our facts need to carry things that are not part of the fact: where it came from, how confident we are, whether a human has reviewed it, and when it was true. Not this document mentions X , but this specific assertion came from page 4 of this version of this document, confidence 0.8, reviewed by nobody yet, valid from March .

Something we've been watching while building N8N Chat UI: US shoppers are starting to buy by asking questions.

Adobe tracked US retail sites over the 2025 holidays. Visits that came from AI assistants like ChatGPT were up more than 7x on the year before, and those visitors converted 31% better than everyone else. They also left less often (33% fewer bounces), and revenue per visit was up 254%.

But it cuts both ways. In a September survey of over 2,000 US adults by Semrush, around 58% of people who use AI said they'd bought something a chatbot recommended, and about the same share said they'd dropped a purchase because of what an AI told them.

Our takeaway: people now arrive at your site used to asking questions and getting a direct answer. If your site can't answer, that question goes to someone else's AI, and you don't control what it says.

Every session of Claude Code or Codex re-reads CLAUDE.md, AGENTS.md, what they import, the rules and the memory index before your first prompt. And pays for it again, every time.

Lancetta's Maintenance pane lists those files, shows what a start costs against the size where Claude Code warns, and runs two dozen checks:

Hey all, let's use this thread for feedback. Thank you!

We've been building a library of ready-made n8n recipes for chatbots, and we're giving them away free. There are 16 so far, and we'd like to know whether we're building the right ones.
What's in there at the moment:

• Lead capture: into HubSpot, Mailchimp, ActiveCampaign, Pipedrive, Airtable or Google Sheets, with consent handled where it matters

• Bookings: Calendly and Cal.com

• Support and orders: Zendesk tickets, plus Shopify and WooCommerce order lookup

• AI: an FAQ bot with email handoff, and a bot that answers from a Supabase knowledge base

• No AI needed: button-menu bots for when a decision tree does the job better and cheaper

They're all at n8nchatui.com/recipes.
What I'd really like to know:

1. What do you end up building over and over for clients?

2. Which tools are we missing? (Notion, Slack, Intercom, Stripe?)

3. Do you prefer AI-driven bots, or simple button flows?

Quick build-in-public update from Sponkly (real-time voice translation for chat, calls and meetings in 32 languages).

This week:

• NVIDIA Inception: accepted

• AssemblyAI Startup Program: accepted

• AWS Activate: $10,000 in credits issued today

What it changed right away: we moved Video Studio's speech recognition to AssemblyAI and raised the video limit from 5 minutes to 1 hour. A 30-minute video is now transcribed in ~10 seconds and is ready with subtitles in under 2 minutes. The AWS credits give us room to scale the backend as usage grows.

At what point does a product become ready to talk to investors?

When you have users?
Revenue?
A clear business model?
A certain number of people using it?

Especially for solo or small-team makers, it feels twice as hard to reach all of these points alone - while continuing to build the product without income and bootstrapping everything.

At some point, it starts to feel like an infinite loop:

A few months ago I posted here that AI access is easy and governance is the next problem. That thread turned into 70+ comments, and honestly, it shaped our roadmap.

The same things kept coming up:

"Governance is the part everyone skips until the bill surprises them"

People using AI outside official tools (shadow AI)

We tested a question asking for research from a particular university, on trapped electrons, within Atomic Physics.

Text search put the correct paper seventh. A basic graph walk put it fifth. Ranking candidates by which relationships actually satisfied the question put it first.

One convenient example proves very little, so we tested 500 questions from the public STaRK-MAG benchmark.

Across two splits, our graph-constraint ranking put a correct paper first on 62.4% and 66.0% of questions. Lexical text search scored 36.4% and 40.4%. A basic graph approach scored 51.2% and 52.4%.

Imagine two buyers asking ChatGPT the same question about the best CRM software.

One manages a startup with five employees and needs affordable software with a simple setup. The other works at a large enterprise that needs advanced reporting, security controls, and integrations with existing systems.

Sponkly for Chrome is live on the Chrome Web Store.

You join Google Meet, Microsoft Teams or Zoom in the browser as usual, click the Sponkly icon, pick the language they speak and yours, and every sentence appears as a live subtitle in your language over the meeting.

• Live subtitles in your language, with the original line underneath

• A translated voice if you would rather listen (the original sound is lowered)

• "Translate my voice too": the others hear you in their language a moment after you speak

• Works on any tab: YouTube, webinars, online classes

• Nobody else has to install anything

• Nothing is recorded or stored

It uses your Sponkly plan's translation minutes, and every new account starts with a 7-day free trial.

I'm trying to get better at separating a useful request from a loud one. A feature can sound urgent in a conversation, but that doesn't tell me how many people need it or whether they'll use it after it ships.

My current approach is to look for the problem behind the request and wait until the same problem appears more than once before it reaches the roadmap. I still worry that this can make a small team too slow to notice a good idea.

What made you decide not to build a feature people kept asking for? Did the data show weak demand, or would the feature have pulled your product away from its main job?

Our WordPress plugin was just approved on WordPress.org.

SpeakBot by Sponkly adds a chat bubble to your WordPress site. Visitors write, call or video call in their own language; your team answers in its own, translated live.

• Translated live chat, plus voice and video calls from the page

• An AI receptionist that answers in the visitor's language when nobody is available, only from the information you give it

• Works with any theme, WooCommerce, the block editor and Elementor

• A loader of about 4 KB, no cookies, no tracking, calls are never recorded

• 32 languages

Setup takes two minutes: install "SpeakBot by Sponkly" from Plugins Add New, create a site in your Sponkly dashboard, and paste the site key into Settings SpeakBot.

A Mac mini or Mac Studio has no built-in screen to fall back on, so when the monitors sleep and drop their link, every window piles up on whichever display comes back first. The write-up covers why it often starts after changing a monitor, a Terminal command to see which display macOS treats as main, and the settings that reduce it, including putting the main monitor on HDMI.
https://manydisplays.com/mac-min...
If you run a Mac mini or Mac Studio with several monitors, I'd like to hear how yours behaves.

Curious where everyone draws the line.

For us it used to be design. Agents could write captions and threads all day, but the moment a post needed a carousel or an image, it was back to a design tool, exporting, uploading, scheduling.

I maintain Termexo, a local-first workbench for coding-agent terminals. v0.11.0 is now available for Windows x64 and both Apple Silicon and Intel Macs.

AI-assisted capability diagram, not an application screenshot. This post was written with AI assistance and checked against the release notes.

I guess most of us started the same way...

One person, one laptop, one agent!

When PH was in its early days, upvotes were the most important. (They are still in demand now), but people increasingly want comments.

Because PH no longer counts only upvotes in the final ranking, but also other things: upvoting, commenting, sharing, etc.

For 3 years I used one productivity tool for everything. Tags, folders, structures, systems. I knew every shortcut. Then one day I opened it and realized I hadn't actually done anything in it for weeks. I'd only been organizing, moving things around, building systems that never got used. I switched to a plain text file, and more work is getting done now.

The tool wasn't bad. I was using it to feel productive instead of being productive.

So for those who've dropped a tool they depended on: was it actually the tool's fault or was it the way you were using it? And the harder question: what tool are you using right now mostly to feel organized not to actually do anything?

AI assistants already work on my Mac. They start dev servers, run builds, and when something is slow, they try to find out why.

All they see is a thousand process names, so they guess. And when they clean up, they can't tell a dev server I forgot last week from the database I'm working in.

This release brings your files to you: a Drawer that slides in over any app, camera cards copied safely to your drives, and every Mail attachment in one place.

A few agents running in parallel...

Each one on its own branch.

Today I have an interview at our national radio station, where I ll be answering questions about myself, RunEvr, and our animation studio.

So let s warm up here first =D

I m a co-founder of RunEvr - an agentic environment for creative teams.

I have 12+ years in animation production, working across almost every stage - from creative work and art direction to a production manager.

Niana remade Wednesday s dance in a T-shirt, blowing up to more than 180M views from her bedroom.

What s another example?

Mine has two phases

My first project was a prototype in lovable, worked great. Then I started adding features such as login and storage/database. Then things started breaking. I started researching and realized I should have

• Planned for authentication from day one

• Connected a database from day one

Phase 2, when I eventually moved it to Claude Code, I should have researched the model capabilities instead of choosing the cheapest model (Haiku ) to build main features. Learned the hard way