Dev Breakfast · 2026-10-01

Someone used a coding agent to generate an entire website and kept getting asked "who designed this site?" He ended up hanging every prototype together with its original prompt on an index page. Whether the same prompt produces good work has nothing to do with the model — it comes down to whether you first pin down a concept, then force it to generate enough versions.

The Same Prompt: Why Some People's Output Looks Like a Designer Made It

First, let's be clear: this is not a tutorial — it's a founder's retrospective on how his website came about. Yakko Majuri, founder of Railcode, generated his whole site with a coding agent, and people kept asking him "who designed this website" and "which designer did you work with." He himself thinks the praise is a bit exaggerated, but because he was asked so many times, he wrote a 10-minute long post pulling every prototype together with the original prompt that generated it into an index page, making the whole process traceable.

Here's the useful part: he didn't ask the agent to "make a nice-looking website" and then pick one. His method was to first pin down a concept — Rollercoaster Tycoon, that roller-coaster game from childhood — and then have the agent generate versions at different intensity levels within the same concept; at least 4 versions per concept, sometimes 10, half of which set direction and the other half he just let the agent run wild with. The wild batch was basically all garbage, but that "neon monster" version is exactly what kicked him off the ASCII plan toward a more ambitious direction. After that came the standard iteration path: first build a playable little game, decide it was too heavy and cut it down to a roller coaster; stick the coaster onto the template and find yourself wondering "what is this doing here," so expand it into a whole theme park; put the park strip on the page and it still looks ugly; in the end, three moves save the day — anchor every asset to the same roller coaster, hand-tune the animations one by one, and change the strip from a standalone module into a color transition leading into the next section. The original article ends here; how it wraps up is never explained.

What's really valuable isn't that the agent can draw — it's that you finally have the ability to describe the picture in your head and iterate on it over and over. The root of design slop has never been the model; it's "you can't articulate what you want." The cost of this workflow is very real: a large share of the versions are doomed to be garbage, and the time goes into picking and cutting rather than generating.

💡 Chef's take: The easiest step to overlook in this method is "store every prototype and its prompt in the index page" — without that record, you'll only remember the last version, and next time you'll start from zero and guess again.

Sources:

• How our vibe coded website looks like a designer made it

Pi.dev: From "No MCP" to Shoving MCP into the Kernel

Pi once proudly displayed a statement on its official site: MCP is not supported. Whenever MCP came up on podcasts, the tone was always dismissive. Upgrade Pi now, and MCP has become a kernel feature — the official blog post is literally titled "You Said No MCP!", picking up the phrase itself and saying it first. Is that embarrassing? Sort of. But what's more interesting is how they justify this reversal.

The reasoning has two layers. One, MCP itself has changed — today's isn't the one from a year ago. Two, they found that the changes required to bring MCP into the kernel are valuable in their own right — for instance, the same set of changes incidentally makes Jev work better in Pi. MCP could have been built as an extension, and someone did build one (pi-mcp-adapter), but in the end it was folded into core. What they're most annoyed about hasn't changed: MCP is hard to compose. Even with a dedicated sandbox like codemode for orchestrating tool calls, it still falls short. The problem is no longer the MCP protocol itself, but the pile of MCP servers out there — many are still written in the old mindset of "cram every tool into context, return plain text to save tokens." Pi's view: MCP should look more like OpenAPI plus smart tool discovery, where tools return structured data and get discovered through documentation and descriptions.

Technically worth noting is codemode. It runs on the harness side rather than in bash's less trustworthy sandbox, so its state goes into the session record, not the filesystem. At its core it's a JavaScript sandbox used to orchestrate and compose tool calls; small JS can be compiled into a WASM binary, and the isolation is decent. In Pi, configure MCP and codemode loads automatically; you can also just make it the default tool — the official wording is "let pi reconfigure itself to enable codemode." Once it's on, what you can do goes beyond MCP — for example, have it call Jev inside codemode, splice Linear MCP together with Jev, and find the 20 most irate commenters in the issue tracker, all without eating into the context.

For people who write code, there's a judgment call here you can steal outright: don't rush to ask "should we integrate MCP?" — first ask "how do the tools get composed?" Protocol support is only an entry ticket; what really determines the experience is whether the server returns structured data or a blob of text, and whether tools can be discovered by their description. Pi's reversal this time isn't a surrender — it's that Pi wants to get into that room and help define the rules. Standing on the sidelines, the rules will never grow in the direction you hope.

Sources:

• hackernews · yarapavan

Gemini 4 Argon Is Priced at $2: Cybersecurity Teams Get It First

Google has released its new flagship model, Gemini 4 Argon: $2 per million input tokens, $10 per million output tokens, with cached input at a 95% discount on the input price. But for now it's only available to a select group of cybersecurity defenders through the Fairwind Program, with developers and enterprises to follow later. The official release says it's already been used internally for C/C++ to Rust migrations and data-center memory optimization, and that its quantum subroutines run 40% faster than the public baseline. Both the pricing and the migration cases are quite specific — but all these numbers come from internal Google scenarios, and external developers can't touch the model yet. A model's capability ceiling is determined by its data; when it truly opens up, first see how it performs on public benchmarks and in your codebase, then talk about whether it's worth it.

Sources:

• hackernews · bradleyg223

Gemini 4 Argon Shows Up on Benchmark Pages, With Every Number Marked "Not Publicly Available"

The last time we discussed it, the new flagship had just made a debut on HN, with performance and pricing completely absent. Today the model page is indeed up on Artificial Analysis, but flip to Intelligence Index, capability index, benchmarks, token usage, cost, pricing, context window — every column reads Not publicly available. Only one thing can be confirmed from the page: it will be run in those 10 evaluations of Intelligence Index v4.3.2, including Terminal-Bench 4.0, SciCode, and Humanity's Last Exam. So for now there are no numbers to compare against, and no price to compare either. Don't rush to treat the chart buzz as a conclusion — wait until it hands over the data before judging.

Sources:

• Gemini 4 Argon (high) - Intelligence, Performance & Price Analysis | Artificial Analysis

The Official MCP Servers Repo Topped the Trending Chart, but It Only Contains 7 Reference Implementations

In our previous issue we just discussed MCP's two faces: on one side, concrete clients integrating it one after another; on the other, teams explicitly saying no. Today the official servers repo reached the GitHub trending chart, which looks like a third signal — the ecosystem is still expanding. But click into it and the repo currently maintains only 7 reference implementations: Filesystem, Git, Fetch, Memory, Sequential Thinking, Time, Everything. Another batch (AWS KB Retrieval, Brave Search, etc.) has been archived and moved out, and the top of the README specifically points to the MCP Registry, saying go there to find server lists. The official positioning of this code is "reference implementations," meant to demonstrate SDK usage, not production-ready; security requirements must be assessed yourself against your own threat model. The repo's popularity shows attention and learning demand, not actual adoption — if you want to get hands-on, reading these implementations to understand how the protocol runs is far more useful than counting stars.

Sources:

• modelcontextprotocol/servers

---

The work in front of you: pin down the concept first and have the agent crank out ten versions, or just throw out a one-liner — "make something pretty"? See you at 8 AM tomorrow.

This issue picked 5 items out of 52 pieces of information across X / Hacker News / GitHub Trending over the past 24 hours (collected and written up hourly throughout the day, fact-checked, then selected and compiled in the morning). Content is generated with LLM assistance; each item is accompanied by a link to its original source — cross-verify any important decisions.