For the previous decade and a half, I’ve been exploring the intersection of expertise, schooling, and design as a professor of cognitive science and design at UC San Diego. A few of you might need learn my latest piece for O’Reilly Radar the place I detailed my journey including AI chat capabilities to Python Tutor, the free visualization device that’s helped hundreds of thousands of programming college students perceive how code executes. That have bought me occupied with my evolving relationship with generative AI as each a device and a collaborator.
I’ve been intrigued by this rising follow known as “vibe coding,” a time period coined by Andrej Karpathy that’s been making waves in tech circles. Simon Willison describes it completely: “After I speak about vibe coding I imply constructing software program with an LLM with out reviewing the code it writes.” The idea is each liberating and barely terrifying—you describe what you want, the AI generates the code, and also you merely run it with out scrutinizing every line, trusting the general “vibe” of what’s been created.
Be taught sooner. Dig deeper. See farther.
My relationship with this method has advanced significantly. In my early days of utilizing AI coding assistants, I used to be that one that meticulously reviewed each single line, typically rewriting vital parts. However as these instruments have improved, I’ve discovered myself regularly letting go of the steering wheel in sure contexts. But I couldn’t totally embrace the pure “vibe coding” philosophy; the professor in me wanted some high quality assurance. This led me to develop what I’ve come to name “vibe checks”—strategic verification factors that present confidence with out reverting to line-by-line code critiques. It’s a center path that’s labored surprisingly effectively for my private tasks, and at present I wish to share some insights from that journey.
Vibe Coding in Apply: Changing 250 HTML Recordsdata to Markdown
I’ve discovered myself more and more turning to vibe coding for these one-off scripts that remedy particular issues in my workflow. These are sometimes duties the place explaining my intent is definitely simpler than writing the code myself, particularly for knowledge processing or file manipulation jobs the place I can simply confirm the outcomes.
Let me stroll you thru a latest instance that completely illustrates this method. For a category I train, I had college students submit responses to a survey utilizing a proprietary internet app that supplied an HTML export choice. This left me with 250 HTML recordsdata containing beneficial scholar suggestions, but it surely was buried in a multitude of pointless markup and styling code. What I actually wished was clear Markdown variations that preserved simply the textual content content material, part headers, and—critically—any hyperlinks college students had included of their responses.
Quite than scripting this conversion script myself, I turned to Claude with an easy request: “Write me a Python script that converts these HTML recordsdata to Markdown, preserving textual content, primary formatting, and hyperlinks.” Claude instructed utilizing the BeautifulSoup library (a stable selection) and generated an entire script that will course of all recordsdata in a listing, making a corresponding Markdown file for every HTML supply.
(Looking back, I noticed I in all probability might have used Pandoc for this conversion job. However within the spirit of vibe coding, I simply went with Claude’s suggestion with out overthinking it. A part of the attraction of vibe coding is bypassing that analysis part the place you evaluate totally different approaches—you simply describe what you need and roll with what you get.)
True to the vibe coding philosophy, I didn’t evaluate the generated code line by line. I merely saved it as a Python file, ran it on my listing of 250 HTML recordsdata, and waited to see what occurred. This “run and see” method is what makes vibe coding each liberating and barely nerve-wracking—you’re trusting the AI’s interpretation of your wants with out verifying the implementation particulars.
Belief and Threat in Vibe Coding: Operating Unreviewed Code
The second I hit “run” on that vibe-coded script, I noticed one thing that may make many builders cringe: I used to be executing fully unreviewed code on my precise pc with actual knowledge. In conventional software program growth, this may be thought-about reckless at greatest. However the dynamics of belief really feel totally different with trendy AI instruments like Claude 3.7 Sonnet, which has constructed up a repute for producing moderately protected and practical code.
My rationalization was partly based mostly on the script’s restricted scope. It was simply studying HTML recordsdata and creating new Markdown recordsdata alongside them—not deleting, modifying present recordsdata, or sending knowledge over the community. After all, that’s assuming the code did precisely what I requested and nothing extra! I had no ensures that it didn’t embody some surprising conduct since I hadn’t checked out a single line.
This highlights a belief relationship that’s evolving between builders and AI coding instruments. I’m way more prepared to vibe code with Claude or ChatGPT than I might be with an unknown AI device from some obscure web site. These established instruments have reputations to take care of, and their mother or father corporations have robust incentives to forestall their programs from producing malicious code.
That mentioned, I’d like to see working programs develop a “restricted execution mode” particularly designed for vibe coding eventualities. Think about with the ability to specify: “Run this Python script, however solely permit it to CREATE new recordsdata on this particular listing, forestall it from overwriting present recordsdata, and block web entry.” This light-weight sandboxing would offer peace of thoughts with out sacrificing comfort. (I point out solely limiting writes quite than reads as a result of Python scripts sometimes must learn varied system recordsdata from throughout the filesystem, making learn restrictions impractical.)
Why not simply use VMs, containers, or cloud companies? As a result of for personal-scale tasks, the comfort of working instantly alone machine is difficult to beat. Organising Docker or importing 250 HTML recordsdata to some cloud service introduces friction that defeats the aim of fast, handy vibe coding. What I would like is to take care of that comfort whereas including simply sufficient security guardrails.
Vibe Checks: Easy Scripts to Confirm AI-Generated Code
OK now come the “vibe checks.” As I discussed earlier, the great factor about these private knowledge processing duties is that I can typically get a way of whether or not the script did what I meant simply by inspecting the output. For my HTML-to-Markdown conversion, I might open up a number of of the ensuing Markdown recordsdata and see in the event that they contained the survey responses I anticipated. This guide spot-checking works moderately effectively for 250 recordsdata, however what about 2,500 or 25,000? At that scale, I’d want one thing extra systematic.
That is the place vibe checks come into play. A vibe test is actually an easier script that verifies a primary property of the output out of your vibe-coded script. The important thing right here is that it must be a lot easier than the unique job, making it simpler to confirm its correctness.
For my HTML-to-Markdown conversion challenge, I noticed I might use an easy precept: Markdown recordsdata must be smaller than their HTML counterparts since we’re stripping away all of the tags. But when a Markdown file is dramatically smaller—say, lower than 40% of the unique HTML measurement—that may point out incomplete processing or content material loss.
So I went again to Claude and vibe coded a test script. This script merely:
- Discovered all corresponding HTML/Markdown file pairs
- Calculated the scale ratio for every pair
- Flagged any Markdown file smaller than 40% of its HTML supply
And lo and behold, the vibe test caught a number of recordsdata the place the conversion was incomplete! The unique script had didn’t correctly extract content material from sure HTML buildings. I took these problematic recordsdata, went again to Claude, and had it refine the unique conversion script to deal with these edge instances.
After a couple of iterations of this suggestions loop—convert, test, establish points, refine—I ultimately reached a degree the place there have been no extra suspiciously small Markdown recordsdata (effectively, there have been nonetheless a couple of beneath 40%, however guide inspection confirmed these have been appropriate conversions of HTML recordsdata with unusually excessive markup-to-content ratios).
Now you may moderately ask: “Should you’re vibe coding the vibe test script too, how are you aware that script is appropriate?” Would you want a vibe test in your vibe test? After which a vibe test for that test? Properly, fortunately, this recursive nightmare has a sensible resolution. The vibe test script is often an order of magnitude easier than the unique job—in my case, simply evaluating file sizes quite than parsing advanced HTML. This simplicity made it possible for me to manually evaluate and confirm the vibe test code, even whereas avoiding reviewing the extra advanced authentic script.
After all, my file measurement ratio test isn’t good. It may well’t inform me if the content material was transformed with the right formatting or if all hyperlinks have been preserved accurately. But it surely gave me an inexpensive confidence that no main content material was lacking, which was my major concern.
Vibe Coding + Vibe Checking: A Pragmatic Center Floor
The take-home message right here is easy however highly effective: if you’re vibe coding, at all times construct in vibe checks. Ask your self: “What easier script might confirm the correctness of my essential vibe-coded resolution?” Even an imperfect verification mechanism dramatically will increase your confidence in outcomes from code you by no means truly reviewed.
This method strikes a pleasant steadiness between the velocity and inventive circulate of pure vibe coding and the reliability of extra rigorous software program growth methodologies. Consider vibe checks as light-weight assessments—not the excellent take a look at suites you’d write for manufacturing code, however sufficient verification to catch apparent failures with out disrupting your momentum.
What excites me concerning the future is the potential for AI coding instruments to recommend acceptable vibe checks robotically. Think about if Claude or comparable instruments couldn’t solely generate your requested script but additionally proactively provide: “Right here’s a easy verification script you may wish to run afterward to make sure the whole lot labored as anticipated.” I think if I had particularly requested for this, Claude might have instructed the file measurement comparability test, however having this constructed into the system’s default conduct can be extremely beneficial. I can envision specialised AI coding assistants that function in a semi-autonomous mode—writing code, producing acceptable checks, operating these checks, and involving you solely when human verification is really wanted.
Mix this with the type of sandboxed execution surroundings I discussed earlier, and also you’d have a vibe coding expertise that’s each liberating and reliable—highly effective sufficient for actual work however with guardrails that forestall catastrophic errors.
And now for the meta twist: This complete weblog publish was itself the product of “vibe running a blog.” In the beginning of our collaboration, I uploaded my earlier O’Reilly article,”Utilizing Generative AI to Construct Generative AI” as a reference doc. This gave Claude the chance to investigate my writing fashion, tone, and typical construction—very like how a human collaborator may learn my earlier work earlier than serving to me write one thing new.
As an alternative of writing the whole publish in a single go, I broke it down into sections and supplied Claude with a top level view for every part one by one. For each part, I included key factors I wished to cowl and generally particular phrasings or ideas to incorporate. Claude then expanded these outlines into totally shaped sections written in my voice. After every part was drafted, I reviewed it—my very own model of a “vibe test”—offering suggestions and requesting revisions till it matched what I wished to say and the way I wished to say it.
This iterative, section-by-section method mirrors the vibe coding methodology I’ve mentioned all through this publish. I didn’t want to write down each sentence myself, however I maintained management over the route, messaging, and ultimate approval. The AI dealt with the execution particulars based mostly on my high-level steerage, and I carried out verification checks at strategic factors quite than micromanaging each phrase.
What’s notably attention-grabbing is how this course of demonstrates the identical ideas of belief, verification, and iteration that I advocated for in vibe coding. I trusted Claude to generate content material in my fashion based mostly on my outlines, however I verified every part earlier than transferring to the subsequent. When one thing didn’t fairly match my intent or tone, we iterated till it did. This balanced method—leveraging AI capabilities whereas sustaining human oversight—appears to be the candy spot for collaborative creation, whether or not you’re producing code or content material.
Epilogue: Behind the Scenes with Claude
[Claude speaking]
Trying again at our vibe running a blog experiment, I ought to acknowledge that Philip famous the ultimate product doesn’t totally seize his genuine voice, regardless of having his O’Reilly article as a reference. However in line with the vibe philosophy itself, he selected to not make investments extreme time in infinite refinements—accepting good-enough quite than good.
Working section-by-section with out seeing the complete construction upfront created challenges, just like portray elements of a mural with out seeing the entire design. I initially fell into the entice of copying his define verbatim quite than remodeling it correctly.
This collaboration highlights each the utility and limitations of AI-assisted content material creation. I can approximate writing types and broaden outlines however nonetheless lack the lived expertise that provides human writing its genuine voice. The perfect outcomes got here when Philip supplied clear route and suggestions.
The meta-example completely illustrates the core thesis: Generative AI works greatest when paired with human steerage, discovering the best steadiness between automation and oversight. “Vibe running a blog” has worth for drafts and descriptions, however like “vibe coding,” some type of human verification stays important to make sure the ultimate product really represents what you wish to say.
[Philip speaking so that humans get the final word…for now]
OK, that is the one half that I wrote by hand: My parting thought when studying over this publish is that I’m not pleased with the writing high quality (sorry Claude!), but when it weren’t for an AI device like Claude, I might not have written it within the first place on account of lack of time and vitality. I had sufficient vitality at present to stipulate some tough concepts, then let Claude do the “vibe running a blog” for me, however not sufficient to totally write, edit, and fret over the wording of a full 2,500-word weblog publish all on my own. Thus, similar to with vibe coding, one of many nice joys of “vibe-ing” is that it drastically lowers the activation vitality of getting began on artistic personal-scale prototypes and tinkering-style tasks. To me, that’s fairly inspiring.
