Vibe coding, one year in. Five tradeoffs, five constants.

Andrej Karpathy coined 'vibe coding' on February 2, 2025. By the time you're reading this in May 2026, the term is about fifteen months old, and the tooling stack has crossed at least two thresholds: per-turn diff editing replaced full-app regeneration, and agentic coding assistants started shipping with browser test loops. Collins English Dictionary picked it as Word of the Year for 2025. The 'one year in' window is roughly Feb 2025 to early 2026.

Because every shipped builder has had to pick a number for each tradeoff, and the number is the position. Take iteration vs one-shot. mk0r's anonymous trial sets ANON_TURN_LIMIT = 6 (src/app/api/chat/route.ts:23). That is not an essay, it is a vote: six is the floor below which a curious visitor cannot evaluate the product, because at fewer turns they never close their first iteration loop. The opinion lives in the integer. Every builder you compare to has its own integer for the same question, and the difference between the integers is the difference between the products.

One-shot is not dead, it shrank. The Quick Haiku mode that lives in mk0r exists for one-shot cases: a tip calculator, a meme generator, a static landing tile. It streams HTML, CSS, and JavaScript out of claude-haiku-4-5 in under 30 seconds with no sandbox boot. The honest scope for one-shot is 'small enough to describe in one sentence and you don't care if the result drifts from your sentence.' Anything with multiple screens, state, or a brand voice that has to stay consistent across the app, you will iterate. The trial budget gives you six turns because below that you cannot tell whether iteration works on your problem.

DEFAULT_APP_BUILDER_SYSTEM_PROMPT at src/core/e2b.ts:184-185 reads: 'You are an expert app builder inside an E2B sandbox with a Vite + React + TypeScript + Tailwind CSS v4 project at /app. The dev server is running on port 5173 with HMR. You have Playwright MCP for browser testing. Read your CLAUDE.md files for detailed workflow, coding standards, and memory instructions.' Nothing about color rules, typography, copywriting, memory format, or design constraints lives in the system prompt. All of that lives in 2438 lines of CLAUDE.md (src/core/vm-claude-md.ts) that are written into the VM at boot. The system prompt is the role; the CLAUDE.md is the manual. The lesson from a year of vibe coding is that prompts you tweak per-session belong in the session, and rules you want enforced every time belong in a file the agent reads on startup.

E2B_TIMEOUT_MS = 3_600_000 lives at src/core/e2b.ts:33 (one hour in milliseconds). The number is a forcing function. A year of vibe coding showed two failure modes for prototypes that linger: people start treating them as production code (the 'vibe coding hangover'), and the underlying sandbox accumulates state that the agent did not author. One hour is long enough to run a real demo session and short enough that the next time you open the project, you are explicitly choosing to resume rather than passively inheriting state. There is a reconnect-after-pause path so you do not lose history, but the default is 'this expires.'

FREE_MODEL = 'haiku' lives at src/app/api/chat/model/route.ts:5. The route refuses any other model selection for unauthenticated or unpaid users (lines 17-24). Two reasons: the free tier still has to converge on a working first draft, and haiku is fast enough that a 30-second preview is realistic. The tradeoff is that you cannot test 'would this work better with a bigger model' without upgrading, which is a real cost for users comparing builders. The honest framing is that the free tier optimizes for 'can a first-time visitor reach a working app' rather than 'can a power user benchmark every model.' Those are different products.

How much of the work that used to happen in the system prompt now happens in a file the agent reads after boot. A year ago, system prompts for app builders were thousand-line megaprompts trying to anticipate every coding decision. Today the prompt is one sentence and the rules live in CLAUDE.md inside the VM. The shift moved the agent from 'told what to do once' to 'reads its memory at the start of every turn.' That's the unglamorous structural change behind the more visible ones like per-turn commits and inline diffs.

Mostly unchanged. The set of apps where vibe coding converges fast (simple state, opinionated styling, no compliance surface) grew this year. The set where it does not (regulated data, multi-user real-time, complex auth, anything where the failure mode is 'wrong answer that looks right') did not get smaller. The honest take after one year is that vibe coding is a different on-ramp, not a replacement, and the people who pair it with reading the generated code get further than the people who treat the output as a black box.

Because the route table is the position. /undo, /redo, /revert, /history, and /cancel all live under src/app/api/chat/. /generate, /regenerate, and /restart do not. The only way to throw away a project and start fresh is to provision a new VM, which is intentionally not a one-click action. Once you commit a turn (commitTurn at src/core/e2b.ts:2073 runs git add -A and git commit in the VM after every successful turn), the cheapest move is always to iterate on top of it. A regenerate button would undo a year of learning that iteration beats one-shot for anything non-trivial.