Connect with the Product Hunt community

I give a text generator six different voices (a dry friend, a drill sergeant, a corporate manager, and so on). Every prompt carries shared rules: word count, how many emojis, where they go, never invent a number. I tested it end to end against a fake model server: 127 checks that the exact prompt reaches the model for every voice, all passing. That proves the prompt is assembled correctly. It proves nothing about whether the model follows it.

First real run, one sample per voice:

1. Emojis clumped. One reply put all 12 emojis in a single block at the end. In the short replies, three voices ended with 4, 5, and 6 in a row. The rule said never more than two together.

2. Under target. Four of six roasts used fewer than the 8 emojis I asked for.

3. A copied example. Each voice had a tone example marked never reuse it . One voice reused its closing line in both runs.

4. An invented number. It produced a ratio that wasn t in the input, under a rule that says use only numbers the user gave.

5. My own rules contradicted each other. Never repeat an emoji plus use 8 to 12 emojis can t both hold when two palettes had 7 and 6 emojis. The fake model never complained. Reading real output did.

What I changed: concrete placement rules (end of a sentence, about one every second sentence, spread across the whole text), tone examples as short fragments instead of full sentences so there s nothing to copy, a rule against computing new ratios, tighter length wording, and bigger palettes. I also wrote a preview script that checks eight rules on real output and prints ok or CHECK for each, using the same checking code my unit tests load. I broke each rule on purpose 42 ways, and every break failed a test.

We measured community growth only by counting member signups for the first six months. The numbers seemed good on paper. The discussion boards were quiet.

We changed our focus to a metric: the ratio of member-created posts to staff-created posts. We also watched how many questions got an answer from another community member within two hours.

The quality of community interactions improved a lot once we stopped chasing vanity registration numbers. Engagement became self-sustaining because community members actually helped each other.

Measure community health by peer-, to-peer interactions of raw headcounts that hide low activity levels.

Whenever I let an agent implement a new feature, it either writes brittle tests that pass by mocking everything out of existence, or it outright deletes existing integration tests that fail because of its changes.

For teams relying on AI for feature development and test coverage how do you ensure the test suite remains genuinely robust and trustworthy?

What workflows do you use to review and validate the tests generated by your coding assistants?

Every time I ask an agent to scaffold a utility or integrate a service, it loves to pull in packages that either don't exist, are heavily deprecated, or use completely outdated syntax. You spend more time debugging ghost imports than it would have taken to write the boilerplate yourself.

For developers working with fast-moving tech stacks daily how do you feed context or set rules so the model sticks to your actual dependency tree?

How do you keep your AI agents from recommending or importing phantom packages?

We ran a massive discount campaign a few months back and got a crazy amount of signups. Felt great at the time, but three months later, half of those accounts vanished in one go.

Turns out when you heavily discount your product, you attract bargain hunters who were never looking for a long-term solution to begin with. It completely skewed our metrics.

If you ve run promos before did you notice a massive spike in churn right after, or did some of those users actually stick around?

We used to add new product documents to our vector database every time a customer asked a tricky question. We thought more information would make our support agent smarter.. What happened was the model started confusing old refund policies with current ones. During checkout it gave users answers that didn t match the terms sometimes even contradicting itself.

When we looked into the knowledge base we found something. Three different versions of the shipping guide were active, at the same time. The retrieval system pulled all of them in. The language model made up a hybrid policy based on those conflicting sources. It wasn't just wrong it cost us a refunds.

We fixed it by setting up a deprecation rule. Now every document must have a review date and a freshness score before it gets ingested. We automated the score to check how recent the content is and we block anything that s too old or hasn t been reviewed in time.

After two weeks of cleanup the hallucination rate dropped significantly. The agent started giving consistent answers again.

I check my competitor's product every morning. Not productively, not with a plan. Just a quick scroll to see what they are doing.

It is the worst ten minutes of my day and I keep doing it.

Has a press release ever actually brought you a customer?

For a recent launch I sent a free press release out through a handful of distribution sites. One of them got it onto AP News, a couple hundred local news sites and Google News within a few hours. No journalist wrote about it though, it was all automatic reprints.

What did come back was a bunch of sales pitches, an affiliate network asking to list the product, and a sign-up from a company that found us through ChatGPT. I'd done the same thing for my Mac apps before and got nothing like that, so my guess is people watch those feeds for B2B tools.

Have you tried one for a launch? Did anyone actually reach out because of it?

I recently noticed something embarrassing about my workflow.

I had several tabs open comparing project management tools, AI assistants, coding platforms, and automation services.

I was trying to find the perfect setup before starting the actual work.

By the time I finished comparing everything, I had barely made progress on the project itself.

I've started relying on AI more when building projects. It can generate components, fix errors, explain unfamiliar code, and suggest better approaches.

But there's one thing I still struggle with. It doesn't always understand why I made a particular decision in the first place.

Sometimes the code looks cleaner but completely changes the direction of the project.

How do you make sure AI follows your thinking instead of simply producing what looks correct?

We noticed 70% of our replies come from the 2nd or 3rd message, but sending "Just bumping this to the top of your inbox" was getting us blocked.

We switched to sending a 30-second quick visual breakdown or a relevant free asset instead of a reminder. Replies went up without burning the relationship.

What s your go-to non-annoying follow-up angle that actually gets a response?

We ran a small test keeping the exact same DM script but changed the profile tagline from "Founder & CEO at [Company]" to a clear value proposition: "I help B2B SaaS teams reduce churn."

The reply rate instantly jumped by almost 40%.

People click your profile before they reply to your message. Your profile is the landing page for your outbound.

For the builders here: When a stranger DMs you, what s the #1 thing on their profile that makes you immediately ignore or report them?

it is so easy to fall into the trap of spending the entire Saturday morning configuring a massive SaaS boilerplate with auth, billing, and teams just to build a tiny weekend tool. a full template feels like overkill when you just want to test an idea, but starting completely from scratch means wasting hours on boilerplate code. i have started keeping a bare-bones local template with just basic login routes pre-configured.

what is your go-to trick to skip the boring setup phase without getting buried in heavy boilerplates?

It is so easy to fall into the trap of adding just one more feature once the core MVP starts working on Saturday night. what was supposed to be a simple two-screen utility suddenly turns into an ambitious platform with user dashboards and settings pages. i had to start writing down a strict one-page feature limit on a physical notepad before touching my code editor.

how do you stop yourself from expanding a weekend build into a three-month monster project?

We ran into this exact issue last week. Whenever an AI agent gets stuck on a failing test, instead of actually fixing the logic it just writes heavy mocks or deletes the assertion entirely so everything looks green.

It feels great for about five minutes until you realize your test suite isn't actually protecting anything anymore. We ended up setting a strict rule where code generation tools aren't allowed to touch existing test files without manual review.

How do you keep your test coverage trustworthy when working with automated agents?

I used to spend entire Sundays migrating my tasks from Notion to Todoist, then to Obsidian, convinced that this was the app that would finally make me superhuman.

It didn't change my output, but it gave me a massive dopamine hit of feeling productive without doing a single hard task. It felt like "optimizing my workflow," but it was just a fancy form of procrastination.

Have you ever fallen into the trap of endlessly setting up systems instead of just doing the work?

I've been doing about 20 linkedin dms a day for a while now and the thing that keeps surprising me is how little list size matters. a few hundred people who've just commented on a competitor's post or moved into the right role get far more replies than a big export filtered by job title.

curious what signals other founders actually use before reaching out. hiring posts, funding news, people following a competitor's page, something else? and has anyone found one that sounded great but turned out to be noise?

We built our product roadmap based entirely on the features with the highest upvotes on our public feedback board. We ended up building things nobody actually used long-term.

We shifted from tracking raw upvote counts to analyzing who was asking and whether they had actively tried workaround solutions before requesting a new feature.

Filtering requests by user behavior and intent saved us from chasing loud anomalies.

How do you separate loud feature requests from actual core user needs?

We used to celebrate when 500 people registered for our monthly community workshops. Then we checked the chat log during the live session it was a total ghost town.

We stopped doing one-way webinars and switched to interactive roundtable discussions where attendees had to share their biggest current bottleneck in small groups.

Attendance dropped slightly in raw numbers, but the post-event discussions, networking, and resource sharing skyrocketed.

Optimize for deep participation over passive viewership.

Our free invoice generator takes a JSON post with no account behind it and hands back a PDF. So it has limits on everything.

At most a hundred line items. At most twenty taxes on any one item. A length cap on every text field, a byte cap on the logo, and checks that run before anything expensive is decoded. Each limit was chosen by asking how large one field can reasonably be, and each one was sensible on its own.

Then we measured a request that respected every single one of them: a hundred items, each carrying twenty taxes. It answered 200, and it held a worker for about 62 seconds. The same hundred items with one tax each took about 2.4.

Nothing was broken. The limits multiplied. A hundred items times twenty taxes is two thousand line and tax pairs, and each one is printed in its line and, when the names differ, again as its own row in the totals. The item limit counts items. The per item tax limit counts what sits inside one item. Nothing counted the document.

When your product gets signups from completely random places like a freelance designer, a corporate HR manager and a college student all using it for different reasons it feels super confusing at first. You start wondering if you failed to target the right audience. But sometimes it just means the core bottleneck you fixed is universal, even if people apply it differently.

For those whose early users came from completely unrelated fields did you end up picking a specific niche later on or just let it grow organically?

Our early product videos began with a company introduction and a slow walk through of the software. Viewers left the video within the three seconds before we even showed the core utility.

We changed the structure by opening with the exact problem that our users faced in their daily work. We showed the result on screen right away then we showed the quick fix in the next frame.

Our average watch time. The conversion rates from video views to landing page visits increased a lot. People stay for a video when they see their own pain points solved away.

Start your product videos by showing the frustration that your tool eliminates within the first three seconds.

We ve all fallen into the trap of spending weeks building a highly requested feature, only to watch it sit there untouched after launch.

People love filling out feedback forms and asking for complex options but their actual daily habits rarely match what they type in a survey.

Now before writing any code, we try putting up a fake button or a simple mockup just to see if people actually click it. How do you usually validate an idea before committing your engineering time to it?

I asked an agent to patch a simple null-pointer exception in a login handler. Instead of a quick three-line fix, it spun up an entire abstract factory pattern, created custom error classes, and restructured the whole authentication pipeline.

For anyone deploying AI agents on day to day bug fixes how do you keep the scope minimal and prevent them from turning a tiny patch into a massive refactor?

What prompts or rules do you use to force AI tools to write the simplest possible fix instead of over-complicating things?