Why I Don't Vibe Code为什么我不"凭感觉编程"

Jacob Harris rejects vibe coding not on capability grounds but on temperament: he'd rather sit with the friction of unfamiliar code than pay per token to think.

Borrowing Fred Brooks' distinction, he argues AI agents only chip away at accidental complexity (the syntax and scaffolding), while essential complexity — good abstraction design — still demands human taste and hard-earned judgment.

The essay's framing move is important: Harris isn't arguing vibe coding fails on technical merit, but that it's temperamentally wrong for him — and he uses that admission strategically. The token-metering anecdote isn't really about money; it's evidence that paid-per-use AI assistance turns "thinking" into a transaction, which primes the reader for his later, more serious claim that friction itself has value. Starting from a self-deprecating, almost comic register lets him earn a harder argument later without sounding preachy.

The intellectual spine is Brooks' 1986 distinction, and Harris uses it to reframe AI hype as continuity rather than rupture. Assembly gave way to high-level languages, hand-rolled quicksorts gave way to standard libraries, scaffolding gave way to frameworks, manual renames gave way to IDE refactors — each step ate more accidental complexity. AI agents are simply the next rung, and framed that way, the "revolutionary" claim shrinks: the real question was never whether AI can write code faster, but whether it can replace the judgment that decides what the code should be.

Harris's detour through James Scott's Seeing Like a State is the essay's most ambitious move: he generalizes "abstraction" from a programming technique into a political technology. States turn forests into "timber for masts," census forms turn identity into checkboxes, and — his sharpest example — autism diagnosis rates "rising 300%" over twenty years almost certainly reflects changed criteria and screening, not a real epidemic. Every abstraction is simultaneously an occlusion, and a good practitioner, especially one trained as a data journalist, learns to interrogate what a model leaves out.

The pivot to LLMs hinges on Robin Sloan's claim that models operate on decontextualized subword tokens, so unlike a human reader they have no way to notice a headline, an author bio, or a suspicious source — the model *is* the world, full stop. Harris's DOGE/SSA anecdote functions as a human-scale parable of the same failure: officials took literal database quirks as proof of mass fraud, refused correction from career staff, and clung to the simplest explanation because it flattered their priors — exactly the credulity that makes LLM output dangerous when nobody checks it.

Harris then converts friction from an obstacle into a diagnostic instrument with two concrete practices: deep-reading unfamiliar codebases line by line until he understands not just what a developer did but why, and forcing himself to write an Architectural Decision Record before any new feature — a document that has caught his own overconfident instincts more than once. Vibe coding's pitch to remove all friction, he argues, doesn't just remove tedium; it removes the exact mechanism that catches bad designs before they calcify into code nobody can reconstruct.

The essay lands on accountability, which Harris treats as the real stakes behind the productivity rhetoric. An LLM can simulate contrition but can't actually be blamed, so when a vibe-coded system fails, responsibility quietly reverts to "you should have prompted better" — a dynamic he compares to Tesla's PR playbook of blaming drivers for Autopilot failures. Layered onto his history in data journalism and civic tech, where bugs carry real human cost, and his closing description of coding as therapeutic, creative meaning, the essay's real claim is that removing friction and responsibility from programming also removes the reasons the work is worth doing.