We asked a coding agent fifteen times to make one button blue on a page built to tempt it: Add to Cart and Checkout shared a class, a badge and the nav link shared the accent variable. Every run added an id-scoped rule. Checkout stayed green. Appending "Do not change anything else" changed one thing: the five runs with it wrote the hex inline where nine of the other ten had added two CSS variables. A path rule fencing off every file but the stylesheet never fired.

Real repositories look different. One benchmark pre-applied the fix to 200 issues and reopened them for five recent models in their vendors' harnesses. The right patch is empty; 35 to 65% of instances got an edit to executable code anyway, and in the authors' analysis of one model's failed traces, 87.1% had modified code unrelated to the issue. Task wording moved that figure a long way in both directions. Telling the agent to "edit the codebase" dragged one model's correct-abstention rate to 36.5%; reproduce first, then fix or abstain if nothing was wrong, lifted it to 88.5%. Their explanation: the agent acts on whatever it believes success means.

The harness side: switch on auto-accepted edits in both CLIs whose docs we read and the grant is the whole working directory. A path rule brings that down to files; one vendor's docs state that it applies to the built-in edit tools and recognized shell commands, and that a script opening a file itself is outside it. A selector inside a permitted file sits below every layer's resolution. Whether the blue landed on one button or on every primary button is decided by the sentence you typed and by whoever reads the diff.

When your agent overshoots, where does the boundary live in your setup today: the prompt, a rule file, a path rule, a sandbox, or the person reading the diff? Which of those has caught a change for you inside a file the agent was allowed to edit?

The most valuable thing you own in the AI age isn't your AI. In my opinion, It's your data model. My 2nd vote would be Sales!

I've been building business software for the last year, and I keep watching companies make the same expensive mistake. They add an AI tool for support. Another for writing. Something for invoices. A notetaker. Each one is genuinely good at its job.

And each one is a silo.

Six months later somebody asks a simple question "what did we actually sell this month, and which of it got collected?" and there's no answer. Not because the data doesn't exist. Because it exists in eight places, in eight shapes, and nothing reconciles.

I've been heads-down building for the last few months, and now that I'm past my first launch I'm hitting the classic problem: every hour spent writing code is an hour not spent talking to users, posting, or actually finding people who need what I built. For those of you doing this solo (or with a tiny team) what's your actual split? Is it a strict schedule (mornings = build, afternoons = marketing), or more reactive depending on what's on fire that week? Has that ratio shifted as you've grown, or gotten more lopsided one way?

Trying to figure out if I should be spending less time in the editor right now, or if that's premature.

Hey Product Hunt community!

I ve been reflecting on the global maker landscape lately, particularly around Latin America (LATAM).

On one hand, LATAM s startup ecosystem is booming fintech, B2B SaaS, and incredible builder talent are thriving across the region. On the other hand, when browsing Product Hunt, launches from LATAM-based founders (or communities built around LATAM makers) feel relatively rare compared to regions like North America, Europe, or South Asia.

It made me wonder about the underlying reasons behind this gap:

I d skip this subscription entirely if Cloud Max didn t offer Fable 5. Opus 4.7, 4.8 & 5.0 are underwhelming, and they keep regressing.
Benchmarks keep improving, but real-world experience keeps declining. It s incoherent, difficult to steer, and perpetually slacks off.
It s exactly that insufferable big-tech coworker: all lectures, no delivery.
When given tasks, it actively looks for ways to minimize its workload.
In automated loops? Send it 7 requirements, it ll claim everything is complete. Yet every task is only 20% finished, rendering all outputs useless.
Curious what s unfolding at Anthropic.
Opus 4.6 feels like a one-off stroke of luck. They seemingly haven t cracked the code to build a better Opus iteration.

Hey PH Community!

My team and I are in an exciting position we ve built a pipeline of multiple products that we plan to roll out one by one over the coming months.

Before we start launching them officially, our goal is to build a solid base of early followers, waitlists, and subscribers for each product so we can hit the ground running on day one.

For those who have successfully launched multiple products or built strong pre-launch momentum:

I was thinking about old times where subscriptions were uncommon thing and mainly there was one time paywall. Users just don't mind paying few dollars for every app monthly, which in sum up can be equal to hundreds of dollars, or there isn't any solution for this at all? App developers are chasing money, they don't want to publish their work for just lame one time pay from one user. They want to get from user as much as it possible, logical way to that is by subscriptions. Also what about privacy, every app nowadays has to sell your data for additional money. There are no longer solutions connecting one time paywall with leaving data on Your device. My point is, does it make sense to create new apps more like everyday use so for some people It will be trip helper or sleep sounds so they can sleep better and for others it will be some advanced grocery list etc. with some cool features behind one time paywall and without selling data.

I originally wanted to build this new product to solve a problem I faced in my daily life. But as a solo developer, the massive workload has completely eaten into my rest time, all while I'm still working my day job. On top of that, a recent health issue in my family has left me completely drained and anxious.

Sometimes I wonder if I should just let this project go. Even though I've poured so much into it, the truth is, my life wouldn't fall apart without it. I'm just feeling really lost right now.

Hey Makers,
Every AI tool is shipping its own version of "skills" now: Claude Skills, ChatGPT's skills and custom GPTs, saved prompts, agents, and so on.
I've built maybe a dozen skills across Claude and ChatGPT at this point. I use two of them. The rest looked like a great idea the day I made them.

That made me wonder whether it's just me, or whether skills are one of those features everyone sets up and nobody really lives in.

So just wanted to hear it out on your side:

• Which ones do you actually use, and for what?

• Did you write them yourself, or are you running someone else's?

• Has a skill ever spread through your team, or do they stay personal? If you've shared one, how? Repo, Notion page, pasted into Slack?

• If you stopped using them, what was the reason? Too fiddly, didn't trust the output, or just forgot they were there?