Kindred Spirits

The honest middle ground between 'AI will change everything' and 'it's all hype.' Jasmine coins 'software vision'—the ability to see which problems are actually solvable with code—and explores what happens when you suddenly have that superpower.

Jasmine describes the addictive phase of rapid prototyping she calls 'Claude Code psychosis'—building custom tools like a YouTube transcript converter styled as Windows XP. But she's honest about limits: many problems aren't software-shaped, requiring insight and motivation rather than automation. The piece balances enthusiasm for creative possibility with skepticism about whether AI-assisted coding truly solves fundamental human challenges.

A raw, honest journey through building a first app with AI assistance. The '290 failures' isn't hyperbole—it's documentation of the actual learning process when you're figuring out how to work with AI coding tools.

Rebecca built Where2Eat, a full-stack web app for group restaurant decisions, in one week using 290 iterations across multiple AI tools (Vercel v0, Claude, GPT). She never wrote code directly—instead, she orchestrated AI tools strategically to create a functional platform that reduces coordination time from hours to minutes. The project proves non-engineers can now build real applications by mastering prompts instead of syntax.

TechCrunch's take on the shift from SaaS subscriptions to custom-built solutions. They coined 'micro apps' for exactly what's happening here—people building precisely what they need instead of paying for bloated software.

TechCrunch identifies the emerging 'micro apps' movement—lightweight, purpose-built applications created by ordinary users rather than professional developers. These aren't polished commercial products; they're personal tools built fast for specific needs. The piece positions this as a fundamental shift in software creation: fun, fast, and fleeting applications replacing expensive subscriptions.

The title says it all. Why pay $10/month for a todo app when you can build exactly what you need in a weekend? This captures the mindset shift that makes daybuilds possible.

Roberto replaced multiple paid subscriptions (Wispr Flow, Loom, Typora) by building personal alternatives with AI-assisted development over weekends. His projects lack commercial polish and wouldn't sell—but that's the point. He values autonomy and customization over production quality, arguing that AI coding has made 'apps on demand' viable for personal use. His prediction: most standalone apps will become features, not products.

"Tech made for people who like making things, not tech made for people trying to appease financiers." Anil Dash's manifesto for orchestrating AI coding agents to build ambitious projects independently.

Dash champions 'codeless' development—orchestrating fleets of AI coding bots via plain-English strategic plans, the same way 'serverless' abstracted away server management. His vision is open-source, vendor-agnostic, and explicitly for independent builders: "Build something that somebody has made a horrible proprietary version of, and release it for free." The target audience isn't enterprises—it's product managers, side-project creators, and anyone who wants to build without needing funding or a team.

One of AI's most prominent researchers shares honest notes on flipping from 80% manual coding to 80% agent coding in a matter of weeks. "Easily the biggest change to my basic coding workflow in ~2 decades of programming."

Karpathy went from 80% manual coding in November to 80% agent coding in December 2025. He's now 'mostly programming in English, a bit sheepishly telling the LLM what code to write... in words.' It hurts the ego, he admits, but the power to operate over large code actions is 'too net useful.' His take: LLM agents crossed a threshold of coherence around December 2025 that caused a genuine phase shift in software engineering—and he's bracing for 2026 as 'the year of the slopacolypse.'

The extreme end of the spectrum: "Code must not be written by humans. Code must not be reviewed by humans." A team that discovered AI agents could achieve compounding correctness in long-horizon coding tasks.

StrongDM's Software Factory team threw out traditional development conventions entirely. Their mantra: 'Why am I doing this? The model should be doing this instead.' They replaced test suites with 'scenarios'—end-to-end user stories that agents can't game—and built behavioral clones of third-party services (Okta, Jira, Slack) to validate at scale. Their benchmark for whether you're using AI hard enough: if you haven't spent $1,000 on tokens today per engineer, there's room to improve.

The title is the whole philosophy. When you just want 'the thing' rather than the process of building the thing, AI coding agents are a game changer for side projects.

Bognanni built FastTab, a custom X11 task switcher in Zig, because existing options frustrated him. He describes the collaborative workflow with Claude—spitballing about the problem, getting a spec, having a working prototype in hours. His honest take on the 80/20 split: AI handles most of the work, but without the taste and experience to steer it, you'll end up with 'a half working prototype that breaks in new and exciting ways every time you ask to pls fix.' The key insight: side projects finally get finished when the tedious parts disappear.

"Over a weekend, between board games and time with my kids, Claude and I built a distributed system with Byzantine fault tolerance." That sentence shouldn't be possible, but here we are.

Garner wrote behavioral specifications in Allium, then Claude generated 4,749 lines of Kotlin with 103 passing tests in 50 minutes. Over the next day—64 commits—he added recovery logic, a REST API, Docker Compose, Kafka integration, and load testing. The result: a distributed system with strong consistency and sub-100ms tail latency, built in stolen weekend hours between family time. It's the most technically ambitious 'weekend project' story in this collection.

A short, enthusiastic reflection on what happens when AI handles the tedious parts—error handling, input validation, propagating properties through multiple layers—and you get to focus on architecture.

Weber's workflow: design testable architecture, write the first test to set a pattern, then describe each test case while AI generates the implementations. The repetitive mechanical work—the stuff that made coding feel like typing exercises—is gone. What's left is the creative part: decisions about structure, trade-offs, and design. His reaction to the shift: 'This is incredible.'

Built a macOS app to export iMessage conversations to PDF—19 source files in one session with Claude, then eight rounds of real-usage feedback to get it right.

Vincent needed to export iMessage, SMS, and RCS conversations with full metadata—timestamps, read receipts, reactions, contact names, embedded images. He discovered that recent iMessage data hides text in binary 'attributedBody' fields using Apple's deprecated typedstream format, requiring NSUnarchiver to decode 99,000 messages. No individual bug was complex, but none would have been found through code review alone—only by actually using the app. Eight feedback rounds turned a working prototype into something reliable.