Vibecoding Comprehension Debate

The Vibecoding Comprehension Debate was a multi-day public discussion on X (July 10–15, 2026) about the challenges of understanding AI-generated codebases. It was sparked by Jan_Wilmake's tweet on the evening of July 10 and became his most-engaged public thread of the period, attracting responses from notable developers including Salvatore Sanfilippo (@antirez, creator of Redis), Mario Zechner (@badlogicgames, creator of libGDX), and Matt Pocock (@mattpocockuk, TypeScript educator).^{[1] [2] [3] [4]}

The Original Tweet (July 10)

At 19:13 on Friday July 10, Jan posted what he later revealed was originally a Claude prompt he decided to tweet instead:^[1]

"so before ai coding it would take much longer to build a codebase up to like 100k loc and the dev would know every bit because they had written every line with sweat and tears but now its like peanuts but the problem is nobody understands crap so now we're all learning how to"

The tweet received 18 likes — Jan's most-engaged post of the week — and spawned 30+ replies by Saturday morning.

Day 2: The Debate Explodes (July 11)

Saturday saw Jan engaging with the best responses and revealing the concrete context behind the abstract reflection:^[2]

• @waincarver (40 likes): compared AI code on top of hand-built projects to "archaeological sites where some culture found old ruins and started building on top of them but with much lower quality materials and structure"
• @iwhaleocean: "we automated writing faster than we automated understanding"
• @mattpocockuk (33 likes): argued the challenge is specific to solo devs — for team leads, "reaching 100K across a team is trivial... manage AI, you just need relatively simple delegation skills"

Jan then revealed his actual situation: "i am working with a largely vibecoded 230k loc codebase right now that im taking responsibility of" — referring to the Hyre codebase. He articulated a three-part plan: (1) create prompting habits that increase understanding as a side-effect, (2) create a CI/CD process with AI review that ensures complexity doesn't accumulate, (3) a third item that trailed off unfinished.

He also endorsed @OSalberger's insight that type interfaces are where the architectural decisions concentrate — a practical strategy for maintaining mental models of AI-written code.^[2]

Day 3: Authority Figures Enter (July 12)

Sunday brought two respected voices into the conversation:^[3]

• @antirez (Salvatore Sanfilippo) argued there are two kinds of AI-coded codebases: those where the developer controlled the main ideas and retains a mental big picture, versus those that were brute-force prompted for results
• @badlogicgames (Mario Zechner, 264 likes) added nuance: "control the types and interfaces, the rest usually falls into place well enough. that said, all the latest models still love to introduce shitty abstractions that work against your types and interfaces. you need to beat them into submission"

Jan retweeted the @badlogicgames post, signalling strong alignment with the types-and-interfaces approach.

Also notable: @dakshgup coined "dead GitHub theory" (83 likes) — observing AIs commenting on and liking each other's GitHub activity (screenshot of Greptile + Cursor bots interacting), a surreal glimpse of AI agents operating autonomously in development workflows.

Day 4: Crystallisation and New Ideas (July 13)

Monday saw Jan synthesise the debate into actionable ideas:^[4]

• RT'd @MatijaSosic (Wasp/OpenSaaS founder): "Before AI, learning was baked into coding... Now, it's no longer the case. You can spend a ton of time 'coding' and still not learn a thing."
• Jan's reply added nuance: "the things we're automatically learning are related to the tooling we use, e.g. I just learned about worktrees. However, I learned very little about the way my refactor was done and what it impacts. understanding was a side-effect of coding, but isn't with vibecoding."
• RT'd @ibuildthecloud (Darren Shepherd, creator of Rancher/k3s): "The faster I move with AI, the more detached I feel from the code. How do you mitigate that? It seems to be a fundamental issue."
• Proposed a concrete product idea to @MatijaSosic and @stevekrouse: AI-generated 60-second PR explainer videos — let AI make a short video about every PR explaining what changed, how it works, and why the changes are correct. @MatijaSosic responded enthusiastically: "AI chooses by itself what you should know and teaches it to you."

Emerging Consensus

By the end of the four-day thread, several principles had crystallised:

1. Types and interfaces first — control the type layer; let AI fill in implementation (@OSalberger, @badlogicgames, @antirez)
2. Understanding was a side-effect of coding — vibecoding removes this; understanding must now be explicitly pursued (Jan, @MatijaSosic)
3. AI-generated comprehension artifacts — PR videos, structured explanations, and forced-explanation prompts as mechanisms to restore the understanding loop
4. Human commit messages — moratorium on AI-written change descriptions (Kenton Varda, endorsed by Jan on Jul 8)
5. Parallel agent workflows — running 6–10 agents with one git worktree each, with a dedicated merge manager agent (@GuiBibeau, bookmarked by Jan)

Day 5: Sandbox Joke, Auto-Mode, and Practical Follow-Through (July 14)

Tuesday saw Jan translate the debate insights into practical workflow configuration:^[5]

• Posted the week's catchiest standalone tweet: "if you're in a sandbox and can't do something, just get out." (7 likes) — a double-meaning about Claude's /sandbox command that sparked a multi-reply thread about Claude Code security models
• Clarified that Claude sandbox is the lightest security option (restricts bash commands only, not full OS isolation) and linked to the official Claude docs
• Turned on Claude's auto-mode and publicly flagged that auto-mode and sandbox may conflict: "maybe it can do this because of auto-mode. I just turned that on. These 2 features may be a bit conflicting"
• Configured "allowUnsandboxedCommands": false after @yan_alex_tech revealed the surprising default
• Closed the "AI reading email" thread from the prior week — agreeing with @hello404agent that reading is a "to-do flag, not proof of eyeballs"
• @hot_town (Open SaaS founder) expressed interest in "baking in" the AI PR video idea from July 13 — confirming the product concept has external demand

Day 6: Brain-Melt Solidarity (July 15)

Wednesday brought the emotional capstone of the debate:^[6]

• Jan RT'd @BraedendotTECH's viral thread (426 reposts): "I'm 33 and I think Claude Code is melting my brain. For 6 months straight I've had 5-6 terminals open at once, waiting…" — a raw description of the cognitive toll of sustained parallel agentic development
• The RT resonated deeply with Jan's own experience of the previous week — discovering sandbox mode, configuring parallel agents, and grappling with the comprehension gap in the Hyre codebase
• Liked @initjean's account of giving Codex a prompt, going to sleep, and waking up to a completed $400 job — the logical extreme of the agentic autonomy spectrum
• Liked @svpino's ironic question: "Why is Anthropic hiring Software Engineers? Why aren't they using agents with loops to do all of this?" (2,033 likes) — capturing the industry-wide tension

Significance

The debate was personally significant for Jan in several ways:

• It publicly established his credibility as someone working at the frontier of AI-assisted development — taking responsibility for a real 230k LOC vibecoded codebase, not theorising
• The types-and-interfaces insight became his practical strategy for navigating the Hyre codebase
• The thread preceded his discovery of Claude sandbox mode (also July 13) — a workflow unlock that eliminated the friction of manually approving bash commands, enabling the parallel agent workflow the community was describing
• The AI PR video idea represents a potential future product or open-source tool

See Also

• Hyre
• Jan_Wilmake
• Claude_Code_Routines
• Self-Driving_Repos
• Spec_First_Development