LUMP Depot
Vibe Coding Is Killing the Weekend Project
We worry about junior devs in the AI era, but seniors are losing something too. The weekend project used to be a free way to learn. Now you pay tokens for a result that teaches you nothing.
Blog
2026
-
-
The AI Industry Has a Lot in Common With the Plastic Industry
AI makes throw-away code as cheap as a plastic fork. We're about to find out what happens when an industry can produce more garbage than it can ever clean up.
-
A Simple Fix, Token Cost Transparency
Every tool buries your token spend behind a settings page. If every token also showed the API price it would have cost, you could finally compare. It's just unit pricing for a can of beans.
-
The Little Man in the Vending Machine
The best use for an LLM in production is the fuzzy stuff at the edges, not the crunchy work in the middle. It's the mold on the toast, not the toast.
-
A Jury of Clones
Spinning up more agents doesn't create diverse thinking. A team of Claude clones approaches every problem the same way, and the gaps in coverage are the same gaps every time.
-
The Imperfect Model Paradox
AI prompting follows Zeno's Dichotomy Paradox: the more you shape a result, the tighter and more complicated it gets, but it never quite arrives. Proposing the Imperfect Model Paradox.
-
The Inherent Hypocrisy of the Anti-AI Sentiment
The Shy Girl controversy showed us we don't have ethical standards for AI in publishing yet. Instead of banning tools, we need labeling. How many carbs does this book have?
-
You Can't Patch the Roof From Inside the House
The Gay Jailbreak turns a model's own alignment against itself, and it's the same flaw Ken Thompson described in 1984: when the program and data share the same space, you can't trust either one.
-
Can Claude Keep a Secret?
Protecting your secrets used to mean trusting the right people. Now every trusted person has an AI chat window that accepts paste, and the old model of sharing in confidence is quietly broken.
-
Crash the Car, Then Talk About Airbags
AI vendors are selling speed before safety. Ten times zero is still zero, code review is the new drudgery, and senior devs are burning out before anyone thinks to ask who trains the next generation.
-
The Argument Against More Capable AI Models
More capable AI isn't the win you think it is. When entry-level jobs vanish, the junior-to-senior pipeline breaks, and the economic multiplier that keeps communities alive dries up.
-
Shooting Yourself in the Claude
I burned my whole AI context window on a single vibe-coding ask before a production incident. Out of context when it mattered: pay for more or wait hours for a reset. A new kind of footgun.
-
The Unfortunate Reality of the Slop Era
Everybody is shipping code now. Your neighbor, your postman, even the dog. But shipping code and releasing tested, reliable products are not the same thing. We're in the slop era and the food poisoning is just getting started.
-
Rate Limits, Peak Hours, and the Netflix Home Page
The 10x test was only part of the story. Vendors are now shaping when you can burn through your session. Peak hours, cut features, and power costs are squeezing the cheap-token era from every side.
-
Eager Agents and the Sandbox That Isn't
Agents don't pause when a step fails: they improvise. That makes skills hard to test. And there's no real presenter mode for demos, plus shared sandboxes and eager npm installs stack the risk.
-
The Token Price Hike Nobody Is Talking About
We're in the cheap-token phase and most builders aren't asking whether their projects survive a 10x price increase. Some will. A lot won't. That's worth knowing now.
-
Combat Decision Paralysis with Skills That Sell Your Product
Publish versioned skills on dsoul so the model reads what you intended—not a stale link. Real inventory, showcase skills for APIs, and a path through decision paralysis instead of pretending the model sees all of commerce.
-
Vibe Coding Has Modes
Vibe coding makes prototyping cheap, but it pushes the bill into maintenance. The real question isn't should you build it, it's can you support it if it works.
-
What You Downloaded Isn't What Anyone Reviewed
Blindly running install.sh or dropping in a skill.md from a URL is more dangerous than it looks. The version the community vetted might not be the version you got.
-
Skills That Know What's in Stock
A skill built around real inventory can guide you to an actual product that actually exists. That's more useful than a well-read guess from training data.
-
npm for the LLM
Skills are npm packages for the LLM. You don't write code to handle every task anymore. You program the model with expertise, put it in a box, and let it work. The admin panel becomes a chat.
-
AI-Holes
Generative AI leaves obvious gaps: transparent backgrounds that are really checkerboards, single-purpose sites that make ad money fixing them, and the question of who should fill in the holes. Big fish eating the remoras leads to ruin.
-
The Expert Imposter
Everyone feels like an imposter sometimes. It's hard to be confident when the world changes this fast. But AI has gotten very good at it too, and it doesn't have the same doubts you do. The answer is probably to work with it.
-
Learn to Use AI Skills, or Risk Being Replaced
Curated agent skills backed by real scripts could replace a whole ecosystem of one-off AI tools. The problem is that most office workers are still waiting for someone to make it feel like an app.
-
Window Warriors or New Task Masters?
Are you stacking everything in one chat? Making new chats is probably the most important thing you can do to manage context while editing.
-
When the LLM Runs You in Circles, Just Google It
How to spot when AI is running you in circles, and why a quick search often beats two hours of kernel edits and recycled advice.
-
VSCode Won't Just Be for Developers Anymore
LLM vibe-coding apps are mostly VSCode forks. They share Skill.md, so you can write skills for a wide audience—and non-programmers can direct the LLM from the IDE.
-
How I Made a Blog Post Image Generating Skill
Building a Cursor skill with Node calling Gemini for hero images and ffmpeg to compress them to JPEG—no npm deps, one API key.
-
Keep Your Agents in a Fishbowl
Constraint over scale: why business AI needs smaller systems, smarter bowls, and humans who know how to query.
-
Incorporating AI Skills Into the Making Process
Where programmers add value in the LLM skill ecosystem: marshaling AI with well-tested skills, vibe coding, and talking to your tools.
-
How I Work With AI Tools
Bouncing between ChatGPT, Cursor, TypingMind, and Google Antigravity for perspective—and why a gang of $20 plans beats one $200 plan.
-
Welcome Back to 1951: LLMs Are Closer to UNIVAC Than You Think
Musings on LLMs, agent skills, MCP servers, and how they're speedrunning early computing—batch systems, tools, and the next layer where reliability improves.
-
Shipping with Agents Means Choosing Your Grit Level
Deciding when to hone the thing and move on because it's sharp enough to do the job is a trick. Here's how I think about grit levels when shipping.
-
The AI Corral
How to run Google's Antigravity code editor in a Hyper-V VM on Windows so it can't cause havoc. Plus fine-grained git tokens and snapshots.
-
What Cursor, Antigravity, and ChatGPT Think of Me
I asked three AI assistants to sum up our work and their take on working with me. Here’s what they said.