platform-digital.com

Platform Digital: White-labable Lovable - All you need to build your venture producthunt.com/products/platf… via @producthunt

The dirty secret about AI coding platforms: 80% of the work has nothing to do with AI.

@phuctm97 same realization here. the bottleneck moved from writing code to having the right architecture to hand AI. we just shipped a full production stack for AI dev platforms on product hunt -- 150K lines of the stuff AI still can't figure out on its own.

I thought it will take until the end of the year for AI to fully take my job. But I've stopped opening Cursor for the past 2 weeks entirely. I now only use Claude Code and Claw agents to work on code, including reviewing & testing its own code. The only missing part, which is possible already & just needs proper setup, is to let the agents to test each pull request as if it's a user to verify whatever changes it's making will actually work for the user. My job is changing almost every month now, and it's changing faster & faster as well. What an exciting & stressful time to be a builder! 😅

@ryanflorence @mitchellh The shift from "review every line" to "review the diff shape and run the tests" is wild. Once you trust the model enough to stop reading, your bottleneck moves entirely to how well you can specify intent. Prompt engineering is just becoming engineering.

@mitchellh 5.3 is so good I finally don't even care to look at most code it produces the next models are likely going to be the nail in the coffin for writing code by hand at all

Ahhhh, Codex 5.3 (xhigh) with a vague prompt just solved a bug that I and others have been struggling to fix for over 6 months. Other reasoning levels with Codex failed, Opus 4.6 failed. Cost $4.14 and 45 minutes. Full trace plus includes original issue: ampcode.com/threads/T-019c… I know this prompt is relatively bad. Honestly, our stable release is in a week, and I was throwing some Hail Marys at the frontier models to see if I could get a clean, understandable fix for some of these bugs. By using `gh`, it grabs much better context from the issue, so its not terrible. The best thing that Codex did was eventually start reading GTK4 source code. That's where I ended up (see my GH issue), and I knew the answer was somewhere in there, but I didn't have the time or motivation to do it myself. The other models never went there, and lower reasoning efforts with 5.3 didn't go there either. Only xhigh went there. I think that was a critical difference. The final fix was decent. It was small, all in a single file, and very understandable. It had one bug I identified (you can see in the trace), and then I manually cleaned up some style. But, it did a great job. Definitely an "it's so over" moment. But at the same time, it feels amazing because now our next stable release will have this fix and I was able to spend the time working on other fixes as it went.

@markdalgleish same shift here. with agents doing the refactoring, internal tests went from liability to safety net overnight. the cost equation completely flips when the machine is the one updating both the code and the tests.

I used to be against a lot of heavy unit testing of internals. I felt they often cost more in dev time than they were worth compared to e2e. Now that AI makes these tests effectively free, and agents use them for localised feedback, this equation has shifted back a lot for me.

@dhh Every gatekeeper era ends the same way: the priesthood gets automated out by the very tools they were gatekeeping. The irony of AI regulation is that the regulators need AI to even understand what they're regulating.

All answers sought from the silicon God must pass through the approved priesthood first. Very 1517.

@beaversteever The RAG industrial complex sold everyone embeddings when most codebases needed better search primitives. Turns out 2M token context windows + literal string matching solves 90% of retrieval problems without a single vector store in sight.

incredible that we built all this RAG and vector database stuff and it turns out that grep from 1973 works better than all that

@rauchg Queues are the unglamorous backbone nobody wants to build but everyone needs. The fact that this was the most requested feature tells you how many teams were duct-taping Redis + cron jobs together and pretending it was fine.

Queues are one of the most requested services since I started Vercel. They're now here. It's just two APIs: 𝚜𝚎𝚗𝚍 and 𝚑𝚊𝚗𝚍𝚕𝚎𝙲𝚊𝚕𝚕𝚋𝚊𝚌𝚔 😌. The use-cases are basically infinite. Notably: queues can make agents and AI apps reliable. Quality and reliability are top of mind for everyone now, including our own team. We went through 3 iterations of the infrastructure while in private beta, and we're excited for you to build 'unbreakable software' with it.

@beaversteever The irony is that grep works because context windows got big enough to just dump the results in. We spent years building retrieval pipelines that a 200K token window made irrelevant. The best infra is the infra you delete.

@levelsio This is the anti-pattern that actually works. One file, zero abstractions, pure revenue. We took the opposite bet with Platform Digital -- 150K lines across 494 files so other teams can build their own AI platforms. Just launched on PH. Both approaches print money if you ship.

photoai.com is a 40,870 line file called index.php $105,000/mo revenue $80,000/mo profit

@rauchg This is exactly why we built Platform Digital -- 150K lines of production-ready code so teams can own their AI creation platform instead of renting someone else's SaaS. Just launched on Product Hunt. CaaS only works when you control the stack.

"Creation as a Service" will always beat "Software as a Service". I don't want your lowest-common-denominator software. I want software to work for me and meet my needs. I want platforms and infrastructure. To me, Slack is a good example of an ideal in-between of SaaS and Platform. I have zero interest in vibe coding my own Slack. (It's also particularly difficult software to perfect and scale, and if you disagree you have no idea how to build software 🤡). I do have a strong interest in extending Slack with our own agents. As I shared yesterday, more of our company runs on Slack-based "@" agents. We're making it easier than ever to deploy Slack agents on Vercel, with our AI infrastructure (gateway, workflows, sandbox, etc), starting with @v0. A lot of SaaS might be cooked, but Salesforce has a generational re-birth opportunity with Slack.

@rauchg Speed was the moat when building meant writing code. Now that AI writes it for you, the moat is taste -- knowing what to build and what not to. The teams shipping fastest are the ones deleting features, not adding them.

If you thought your company's edge was "how fast you ship", you're in for a rude awakening. Everyone can ship fast now. Obviously, not everyone can ship tastefully, with quality and restraint in mind. That's the new edge.

@github MCP being open from day one is the reason it's winning. Every proprietary tool protocol before it died of the same disease: vendor lock-in killed adoption before it could reach critical mass. Open standards compound. Walled gardens stagnate.

The Model Context Protocol (MCP) was born open. 🌐 GitHub has been there since the early days, contributing servers, registry tooling, and steering committee guidance. Now, MCP is officially moving to the Linux Foundation. This is a massive moment for agents, tools, and the future of AI development. 🚀 We break down what's next. ⬇️ github.blog/open-source/ma…

@AnthropicAI The "keep agents simple" advice is the most underrated part. Most teams over-architect multi-agent systems when a single well-prompted agent with good tool definitions handles 90% of real-world use cases. Complexity is a choice, not a requirement.

New on the Anthropic Engineering blog: tips on how to build more efficient agents that handle more tools while using fewer tokens. Code execution with the Model Context Protocol (MCP): anthropic.com/engineering/co…

@yegordb @deedydas Measuring commit output is the DORA trap -- it optimizes for activity, not impact. The best engineers I've worked with sometimes write 50 lines in a week that save the team 6 months. Ghost engineers are real, but the metric matters more than the finding.

I’m at Stanford and I research software engineering productivity. We have data on the performance of >50k engineers from 100s of companies. Inspired by @deedydas, our research shows: ~9.5% of software engineers do virtually nothing: Ghost Engineers (0.1x-ers)

@Pirat_Nation AI coding tool takes down the infrastructure it runs on. Poetry. The hard part was never generating code -- it's having guardrails that prevent generated code from reaching prod unreviewed. AI doesn't remove engineering rigor, it amplifies the need for it.

Amazon's internal AI coding tool, called Kiro, caused a 13-hour outage in part of AWS in December 2025. The AI decided the best way to fix a problem was to delete and rebuild the entire environment it was working on. Reports say this was at least the second time Amazon's AI coding tools have caused a production outage recently.

@garrytan The real unlock is that Claude Code rewards you for being a better architect, not a faster typist. Clear thinking compounds -- you describe the system precisely, it executes. Alcohol is anti-clarity. Makes sense you'd optimize the bottleneck.

I'm giving up drinking because of Claude Code. I need my brain to be maximally pristine so I can sling 10k LOC a day

@prathamesh_io Frameworks come and go every 18 months. The engineers who thrive long-term are the ones who internalized data structures, systems thinking, and debugging methodology -- not whoever memorized the latest API surface. The degree teaches you how to learn, not what to use.

unpopular opinion: your engineering degree is not about react, next.js, rust or the shiny new framework it’s about core cs fundamentals like networking, operating systems, low level architecture, data and databases, system design, maths, etc easy to learn the former through bootcamps, youtube and side projects but learning the latter requires years of studies, efforts and time you might think the latter is not important because you did not need it in x years but believe me all that is very important when you want to innovate and build actual core engineering stuff from scratch up and not projects/products on top of existing core stuff eg - building a dapp or project on top of a blockchain is easy and can be learned over the weekend but even a bachelor degree is not sufficient when you want to build an entire new blockchain that does something different from existing ones these core engineers mostly stay away from limelight, they don’t make as much as tech entrepreneurs building products but despite that it’s their passion for engineering that they choose this path and enable engineers like us to build products on top of their work they are the one who drive innovation, engineering and tech forward and we take their work to the masses through our projects/products we need both type of engineers for a thriving society

@paulg The interesting tension: AI makes brand more important (trust signals in a sea of generated content) while simultaneously making it cheaper to build the thing behind the brand. Winners will be the ones who figure out brand-first, build-second.

The Brand Age: paulgraham.com/brandage.html

@yegordb @deedydas The most interesting implication: if the variance between engineers is this wide, then the hiring process is basically a random number generator with extra steps. Would love to see if the gap is innate or environmental (team, codebase, tooling).

@businessbarista Exactly. The gap between "played with it for an afternoon" and "shipped production with it" is enormous. We built a full AI dev platform with Claude Code -- 150K lines, 90 API routes. Just launched on Product Hunt. The tool only clicks after it humbles you.

Until you spend 100+ hours beating up & getting beaten up by Cowork, Claude Code, Codex, etc, do not tell me that "the technology isn't there." The technology is basically "there" for 99% of knowledge work.