Ship a Weekend AI Project With a Failing Build

The build says no before the weekend does

You have a weekend, a browser tab, and an agent that wants to build a game. tibo.dog spent $5.99, the domain purchase. The trick is to make the small version the only version that can finish. If the agent can add a folder, a dependency, or a large sprite, the project has already escaped the weekend.

Lauren Tan, who posts as @poteto, launched tibo's game on Nov. 1 and credited Grok Bot and Cursor with building it. The tibo.dog build ran for four days, producing about 1,150 pull requests. That volume is the warning; the rules are the lesson.

A small browser project is a good test case because the loop is visible: input, state, render, score. Most side projects stay side projects, and that is fine. The weekend is short, so the build has to carry the discipline. The agent can touch each part without a long setup, and the workflow needs a build that rejects growth before it becomes a maintenance project.

The checklist keeps the weekend alive

Run these gates before every commit. Each one is cheap to test: make the build fail when the limit is crossed. You can run the whole list before the coffee goes cold.

Domain-only cost: decide before the first commit that the only paid line is the domain. If a tool asks for a card, stop and find a free tier or local path.
Hard gzip size budget: pick a compressed bundle ceiling and make the build fail when it crosses it. The agent should see the failure, not a warning.
Lint rule per bug: when a bug is fixed, add a rule that catches the same class of mistake. The next agent run should be blocked by the rule, not by memory.
No comments: if code needs an explanation, rename it or add a test. Comments become stale faster than the agent can read them.
No binary assets: keep images, fonts, and generated blobs out of the repo. If the game needs a sprite, generate it from code or reject the asset.
Deterministic replay: make the same inputs produce the same state. A replayable run is easier to debug, easier to verify, and harder to fake.
Headless bot tuning: let a simple bot play in a headless browser before you polish the UI. The bot exposes broken loops, stuck states, and unfair difficulty.

The tibo.dog rules show how strict a small build can be.

The size gate is the most visible one. tibo.dog's release lands near 124 KB after gzip, and the build reportedly fails when that budget is crossed. tibo.dog uses linting to ban comments, forcing better names or tests. tibo.dog puts frame-by-frame logic in .hot.ts files, and those files forbid memory allocation.

Replay makes a leaderboard honest. Because the simulation is deterministic, the server can re-run leaderboard submissions and catch cheating. SonarQube, Semgrep, ESLint, Pylint, and RuboCop are examples of static analyzers that can become repeatable compliance gates when build inputs, rule configuration, and analyzer version are pinned. One scar-based workflow replaced advisory prose with four concrete engineering mechanisms, then reported a 0.0% repeat regression rate over 238 production tickets and over 588 codified SCAR entries.

Agent volume becomes a test when gates are fixed

Claude Code guidance keeps instruction files under 200 lines, since longer files use more context and get followed less reliably. Anthropic's Applied AI team said good context engineering means picking the smallest high-signal token set that makes a desired outcome likely, because each input token uses a finite attention budget. Claude Code can block tools, commands, or paths through enforced settings and hooks, even when the model would choose them; hooks can also stop an action outright at fixed lifecycle points.

Greptile's free Starter plan gives one active developer 50 credits per month and unlimited repositories. Anthropic said its typical engineer merged 8 times the code volume per day in 2026's second quarter compared with 2024. If the build fails, fix the rule before you ask the agent for another patch.