Connect with the Product Hunt community

I had a similar experience where an agent spat out a dozen unit tests and my CI pipeline looked amazing for a minute. Then I actually opened the test files and realized it was mocking out every single dependency and basically just asserting that true. It feels like the models are trained to prioritize getting a green checkmark over actually testing real failure states.

Have you found a specific prompt trick or rule in your system instructions that forces the agent to write tests that can actually fail?

I used to think that giving an AI assistant access to every single file in my project folder was the ultimate cheat code. But lately I ve noticed the exact opposite happening.

When it sees old legacy code from three years ago alongside my clean new components, it gets confused and starts copying the bad habits of the codebase. Sometimes a smaller tightly controlled prompt works way better than letting it scan everything.

Have you started restricting your coding assistant's file access, or do you still feed it the whole repo?

Parsing messy public web data is honestly harder than building the actual core feature.

You think a website's meta tags will have standard company info and then you find sites that put random blog text in the title tag. If your tool makes a wild guess and gets it wrong the user instantly loses trust in the whole product. I've started marking scraped fields as "needs review" whenever the confidence score drops below a certain point.

Have you found a reliable way to clean up dirty data before it hits the user interface?

It s fun setting up AI agents to handle routine backend tasks until you realize how easily a weird input can mess things up.

Even with good guardrails giving an agent direct write access to a live database always feels like playing with fire. I've started forcing a manual approval step for anything that modifies or deletes existing records no matter how confident the script claims it is.

Where do you personally draw the line do you let agents touch production data automatically or do you keep a human check on every single write?

People rarely type out a long paragraph explaining every little annoyance when they cancel a subscription they usually just check a generic box like "missing feature" or "too expensive" to get it over with. But when you actually look at session replays you notice users getting stuck on a weird button placement or a confusing login screen for three minutes straight before dropping off. Those tiny micro-frustrations add up yet nobody ever puts them in a feedback form.

What methods do you use to spot these hidden drop-off points?

When you're starting out it feels natural to build everything yourself internal docs task trackers chat systems just to keep users inside your app.

But at some point you realize users literally hate logging into yet another dashboard just to check one notification. They want your product to meet them where they already spend their entire day. It s super painful to put a pause on your own product roadmap to spend weeks building external integrations.

How did you figure out the exact moment to shift focus from native features to ecosystem plugins?

The Sunday Roundup from Product Hunt began with the news about Reddit hoping that users would sign up, and it reminded me of all the times I've seen my product searches followed by the word reddit. It's something I add on purpose, since I'm looking for rough feedback from users who actually purchased the product, rather than from a review website that earns a commission.

If there is more of such content locked behind logins, then I'll be reading fewer reviews. I'm guessing others would do the same. This leaves an opening for those who need to know whether a particular tool is worth spending money on.

There's this weird pressure in tech where being available 24/7 is treated like the ultimate badge of honor. I tried stepping back from answering every random comment and non-urgent email right away, and honestly, half the people just went ahead and figured the solution out on their own. It made me realize how much time I was wasting babysitting threads that didn't need me. Have you ever tried going radio silent for a few hours, and did anything actually break when you did?

I've started asking one thing before I give anyone advice about their product. Do you want me to help you think, or do you want me to tell you what I would do?

They are completely different conversations and most of the time people want one and receive the other.

Ask someone what they think of your product and you get a polite opinion with nothing underneath it.

Ask them what they did last time they used it and you get a story. A specific moment. The thing that broke, the thing they skipped, the thing that surprised them.

One thing I've learned while building software is that there's a difference between successfully deploying code and knowing the application is ready for real users.

A deployment can complete without errors, the application can load, and the main features can appear to work.

But that doesn't necessarily mean everything is functioning correctly.

For example:

been testing multi-agent setups with tools like cursor and claudecode recently. they were great early on, but giving them more autonomy seems to backfire.

instead of sticking to the main task, they end up hunting down tiny bugs or adding unnecessary code tweaks that just drain hours in loops.

is this just a prompt control issue or are agents hitting a limit with complex context?

Version one. Version two. Version three. Version four.

Each one felt better than the last. Each one converted about the same.

I used to keep my chat apps open all day replying to every ping within five minutes because I thought it showed dedication.

All it really did was chop my day into tiny useless pieces where I couldn't focus on anything meaningful. The day I started batching my replies to just twice a day my anxiety dropped and I actually finished things on time. Nobody got mad and the world didn't end.

How do you manage your communication loops without letting constant notifications wreck your focus?

I've been using AI for the last 3 months to generate images, code snippets, and ideas for small projects. It's super fast, but I noticed I m starting to prompt instead of thinking from scratch, and my own ideas feel less original.

For those of you who build products with AI tools every day has AI made you personally more creative or less creative? Do you have any rule you follow to make sure you still think for yourself and not just copy what AI gives you?

It's so tempting to keep saying yes when a client is writing you big checks every month even if they treat your team like an in house agency instead of a product company.

A couple of years ago, we had one major account bringing in almost half our revenue, but every single week was total chaos with custom feature demands. Letting them go felt like jumping off a cliff financially but suddenly our engineering bandwidth opened up for actual product improvements.

If you've ever walked away from a massive account, how long did it take for your MRR to bounce back?

It's funny how smart these models are, yet they will completely ignore the clean patterns right in front of them and mimic some weird workaround I wrote in a utility file two years ago.

The more context or files I attach the more it hallucinates random functions that don't even make sense for what I'm building today.

I'm starting to think manual context curation is the only way to keep the code clean. How do you feed your tools the right files without overwhelming them with noise?

It's honestly hilarious and frustrating at the same time. I asked my coding assistant to fix a tiny typo in a button text yesterday and it came back with a brand new component architecture split files and a custom state hook. It felt like dealing with an overly enthusiastic junior developer high on espresso who just wants to show off everything they can do. I just wanted a one line string change!

Have you found a reliable way to keep these agents lazy and focused strictly on the exact line you point to?

Two years ago I sat down and wrote out where this business would be in five years. Revenue targets, team size, the whole picture. I was specific. I was confident.

I found that file last week and read it start to finish, and my first instinct was to close it and pretend I had not seen it.

I ran into this exact headache while pulling company details from random websites.

Some pages have their logo hidden in weird spots or completely skip a clear description, leaving my script scrambling. At first, I tried forcing smart guesses just so the UI wouldn't look empty but it backfired when the tool confidently displayed the wrong info to users.

Now I'd rather just leave a clean "not found" or let the user fill it in manually than show fake confidence. How do you handle UI states when your scraper comes back empty handed?

It's super when customers quietly slip away and you have no idea why because everything looks fine on paper. Users hate filling out long feedback forms so they won't email support just to say a button layout is annoying they'll just quietly switch to something else. Hopping on quick user calls or looking at session replays always shows the tiny blind spots you completely missed.

What is your go to method for spotting those silent usability issues before people churn?

Having a solid component library saves a ton of time but sometimes it feels like it takes away the freedom to actually think about what a specific screen needs. You end up trying to cram a unique feature into an old component box because "that's how we always do it.

" Products usually feel much better when the core experience stays coherent but individual screens are allowed to solve their specific problems naturally.

Do you prefer strict design guidelines, or do you give your team the freedom to invent new patterns when needed?

I let an agent auto-generate unit tests for a complex module. It created 15 test cases, all green! But when I actually looked at the assertions, it was mocking everything into oblivion and testing getters and setters. How do you prompt agents to write rigorous, edge-case-heavy tests instead of vanity metrics?

I watched a stranger use our product for the first time last week. Screen share, I just stayed quiet and watched.

It was painful

She skipped the welcome screen, missed the main button, and spent almost two minutes in settings looking for something that was right on the dashboard.

Everything we thought was obvious wasn't.

You don t really read it, and now it has to be reviewed by your colleagues, and they are like, this doesn t look right,'
You re just letting AI do the work for you.
The words of Shopify's CEO.
So how do you tackle it? Anything that is working on your end?