Telling Claude to "make it better" over and over does not converge on a better result. In a widely shared experiment, one creator ran that exact prompt 100 times on a coffee shop website. By their account, the useful upgrades stopped around round 32, and everything after that was bloat. The fix is not a smarter loop prompt. It is a loop with a concrete success criterion, a saved original, and a human who inspects each round.
The experiment: "make it better" 100 times
Nate Herk, a YouTube creator covering AI automation, told Claude Opus to make a simple coffee shop website better 100 consecutive times, with no other direction. We could not independently reproduce or confirm the details, so treat this as one creator's experiment, not a controlled study.
The pattern he reported is the useful part. Early rounds added real value, including working online order forms and some background animations. Past about round 32, the additions stopped serving the site. By round 100 the layout was broken, the navigation was over-engineered, and a barely functioning virtual pumpkin carving tool had appeared.
Why a vague command turns into bloat
An AI model is built to fulfil the prompt in front of it. "Make it better" gives no definition of better and no point at which to stop, so the easiest way to respond is to add something. Addition looks like improvement, even when it is not.
This is not a new observation. Max Woolf ran a similar experiment asking a model to keep writing better code, and Simon Willison covered it: later rounds sometimes added bugs and enterprise-style patterns nobody asked for. Claude prompting guides say much the same thing. "Make it better" and "try again" add no new information, so the model guesses the way it did the first time.
The building comparison from the episode fits. You want a blueprint first and an inspection before anyone moves in. A loop with no inspection phase is construction with no inspector.
If you use Claude Code, the command line tool that runs in a text terminal, long loops carry a second cost. Claude Code uses sub-agents, small background assistants that double check the main model's work, and they summarize their status on a timer. To write a quick status label, the system feeds the conversation history back into the model.
By one creator's measurement, that background activity can silently drain up to 20% of your usage limits. That figure comes from a single creator and has not been confirmed by Anthropic. If it holds up for you, an open-ended loop is paying for the AI to narrate its own bloated process.
What works instead: a loop with a fixed judge
The better pattern is a short loop where something fixed decides whether a round helped.
Three specific prompts beat 100 vague ones. One creator says they built a fully functional Roblox game, with stages, boss fights, and unique character auras, using three highly structured prompts. They defined the scope of each prompt, set strict boundaries on what the AI should write, and paused to evaluate the code after every step.
Research points the same way. The Self-Refine paper (draft, self-critique, revise) reports roughly a 20% average absolute improvement, and notes that the gain depends heavily on how strong the base model is. A GitHub issue in the ProbOS repo, drafted by a Claude Code agent, states the discipline plainly: a revision replaces its parent only after a fixed evaluator says it is better.
Loops suit checkable tasks. Anthropic's ralph-wiggum plugin reruns a task until it is done, and write-ups say these loops work best on deterministic jobs with a clear pass/fail check, like "all tests pass." They struggle on subjective work such as UI aesthetics, which is exactly what a coffee shop website is.
Long correction chains rot context. A Business Today piece describes how a long trail of corrections builds up in the conversation. Its fix is to go back to the original prompt, improve it, and regenerate.
Three things to do today
Replace "make it better" with one concrete success criterion. Tell Claude what done looks like, such as "the order form works on mobile," so it knows when to stop.
Save your original before asking for a revision. That gives you a clean rollback point before any human review.
In Claude Code, check your sub-agent settings. One creator says setting
CLAUDE_CODE_FORK_SUBAGENTS=0stops background agents from eating your usage limits. Test it on your own setup before relying on it.
The bigger point: taste is the asset
The instinct to let the machine do the heavy lifting is strong. But relying entirely on an AI to define what "better" means is outsourcing your own taste, and taste is the one asset the model cannot replicate for you. Use the AI to draft and assemble. Keep the blueprint and the inspection in your own hands.
FAQ
What happens if you tell Claude "make it better" 100 times?
According to one creator's experiment on a coffee shop website, early rounds added real improvements, but after roughly round 32 the additions became bloat, and by round 100 the site's layout was broken. Treat it as one data point, not a benchmark.
Why does a vague "make it better" prompt make results worse?
The prompt gives the model no definition of better and no stopping point, so adding features is the easiest way to respond. Without a concrete criterion, addition gets mistaken for improvement.
How do I run a safe revision loop with Claude?
Give it one concrete success criterion, save your original version first, keep each prompt narrowly scoped, and review the result after every round before continuing.
Can background sub-agents use up my Claude Code limits?
One creator measured background status summaries consuming up to 20% of usage limits in Claude Code. That is a single, unconfirmed measurement, so check your own usage before changing settings.
Companion episode:
Want more practical AI breakdowns like this one? Head to joebuildsai.com.
