VibeCoding is writing what you want and magic happens - kind of. Great for party, useless for production. So, in my company many people have start to write detailed Specs. Still the same process, just more "professional" - they thought. The crazy thing is, that they did the SPECs for the AI manually! How stupid! They have even followed a course and have bought a book!!!! about how to write Specs for vibecoding. Then I showed them mine I have got GENERATED for a simple application and they were blown away. So the point and paradigm is the following: Writing more detailed prompts does not help when apps are getting complex - WHEN you start vibcoding from there! What you have to do is - from the same abstract prompt - generate the VISION.md, SPEC.md, ... your platform needs for context. Never write those files yourself! Always generate them from your initial App description. Then iterate over the GENERATED Specs and when things are wrong or missing tell you AI to fix it - but never CHANGE those context files by yourself! The funny thing is, that for non-engineers that has always been the case. None of them would have dreamt of writing Specs by themselves. Its only engineers doing that. They really have to unlearn manual Specs writing! To give you an example how a modern SPEC.md looks -> check https://github.com/realB12/ApplAi_Generator/blob/main/VSC/client/AI/KIMI/CONFIG/SPEC.md

When AI output doesn't sound like you, the instinct is to explain harder. You add more adjectives. Professional but friendly. Confident but not salesy. Warm, but concise. And the output still comes back not quite right, because words describing a style are a poor substitute for the style itself. Think about how this plays out. You want AI to draft something in your voice: a client update, a proposal, the kind of thing you write every week. So you describe how you want it to sound. You type out "keep it clear and warm and professional," maybe throw in a few more instructions, and hit send. What comes back is generic. It's technically what you asked for, but it reads like anyone wrote it. So you rewrite it by hand until it sounds like you again. Next week, same task, same vague description, same disappointing draft, same manual rewrite. You're describing your style over and over and never quite landing it. ---------- THE REAL PROBLEM ---------- The problem is not "AI can't match my voice." The problem is "I keep describing my style in words when I should be showing it real examples." Your style isn't really capturable in a list of adjectives. The way you open a client email, how long your paragraphs run, where you land between formal and casual, the phrases you reach for: that lives in the writing you've already done, not in a description of it. When you only describe it, the AI has to guess at all the specifics, and it guesses generically. That's not a capability gap. It's an input gap. You're handing it a description when you could hand it the real thing. ---------- WHY THIS MATTERS ---------- Describing your style instead of showing it costs you twice. First, the output is never quite usable, so you rewrite every draft by hand. The AI saves you less time than it should, because you're always doing the last mile yourself to put your voice back in. Second, it never gets better. Because you're describing from scratch each time, in slightly different words, the results stay inconsistent. One day it's close, the next it's off, and you can't build any reliability on top of it. You start to think AI just can't do your kind of writing, when the truth is it was never actually shown what your kind of writing looks like.

On Tueday, Anthropic announced invisible watermarks in Claude's output. Within 24 hours, a "watermarks-remover" repo was live on GitHub. Version 0.3.0, CI passing, MIT license, and it doesn't stop at Claude: it targets Gemini's SynthID, OpenAI's provenance surfaces, and C2PA file metadata in one package, framed politely as "hygiene on content you own." I'd call this the "token washing economy": every marking scheme creates a market for unmarking, and the economics run entirely one-way. Marking text costs the labs engineering effort, legal review, and EU coordination. Unmarking it costs one rewrite pass through any other model (the statistical signature doesn't survive paraphrasing, which Anthropic's own documentation concedes). Three things follow for anyone building on or around this: → Provenance marks are evidence for cooperative actors only. They'll catch the lazy and the honest, never the motivated. → Any workflow that treats a watermark check as a control (hiring, publishing, education) is now auditing for effort, not authorship. → The compliance layer and the detection layer are officially separate markets. The first is mandatory and works. The second is voluntary and doesn't. None of this makes the EU AI Act's transparency rules pointless. It makes them what they always were: disclosure obligations for the labs, not truth machines for the rest of us.

As candidates increasingly use AI to prepare for interviews, refine their applications, and in some cases even assist in real time during remote interviews, traditional hiring processes are struggling to actually assess what they were originally designed to assess. This is a specific and costly consequence of AI adoption showing up disproportionately on one side of the hiring table, the candidate side, without a corresponding adjustment on the employer side, and businesses that haven't adapted are spending more time interviewing while getting worse signal than before. ------------- Context ------------- Traditional interview processes were designed with a set of implicit assumptions about what a candidate's answers, on the spot, in real time, actually revealed about their genuine knowledge, judgment, and communication ability. AI has quietly undermined several of these assumptions. Candidates can now use AI extensively to prepare polished answers to common interview questions in advance, and in some remote interview settings, there have been documented instances of candidates receiving real-time AI assistance during the interview itself, generating responses that the candidate then relays, without the interviewer necessarily being aware this is happening. The result is a specific degradation in interview signal quality: the traditional interview format increasingly measures a candidate's ability to prepare with or use AI assistance well, rather than measuring the underlying knowledge, judgment, and communication skill the interview was originally designed to assess. Businesses that continue running interviews exactly as they did before AI became widely accessible are, in a meaningful number of cases, making hiring decisions based on signal that's considerably less reliable than it used to be, without necessarily realizing the reliability has degraded. ------------- Where This Degradation Becomes Visible ------------- A small company's hiring manager described discovering this gap directly after a series of hires that hadn't worked out as well as their strong interview performances had suggested. Digging into what had gone wrong, she began to suspect, based on some specific patterns in interview responses, that at least some of the candidates had been using AI assistance more extensively than the traditional interview format was designed to account for, whether through extensive advance preparation that produced unusually polished but somewhat generic answers, or through less certain but plausible real-time assistance during remote interviews.

Preparing for my next vibcoding adventure, currently feels like preparing for some adventurous journey through the dark where you will go through without really noticing that you are going through it, but with something you will have got at the end, that might change the way you see life and yourself. It is some kind of mystical process. Preparing the journey by filling the VISION.md and SPEC.md is holy. It's a ritual. It's pure as it is humiliating. Different from packing our backpack, where the journey and the outcome is defined, packing for VibeCoding feels different: vague by nature, there is no urge to be complete or perfect when the final destination is something between wonder, surprise, and a total mess. Vibprepping feels like the Holy Spirit copiloting and taking care of your incompetence and mildly nodding at your humble trying the best you can from knowing nothing. VibeCoding has become a spacebending arc-flight experience that, beyond delivering hands-on, substantial, tangible artifacts, comes with some subtle, transformative, supernatural, ghost-like, permanent side effects echoing in my soul: as if some previously learned esoteric manifestation gimmicks have become an everlasting hardcore reality - kind of. Ready with all the context and AI preparation files, hovering over the Enter button to unleash an energy-devouring time and substance-bending global machinery to my service feels like a climax in a religious transformational ritual where all spirits concentrate in a millisecond touch to give birth to something with truth and eternity - a holy sensation that might have inspired Michelangelo for his "Creation of Adam".