Loop Engineering Is Replacing Prompts (How i'm using it)

Loop syntax replaces hand-written prompts in Cursor

Stop saving prompt lists. The creator's new habit is a single one-shot prompt that ends with a loop instruction (rephrased as loop until it's done and until you've tested play through it end to end since Cursor has no native /loop command). Combined with Fable 5, this collapses what used to require a PRD and the BMad method into a single requirement dump.

The Yeats → 3D game build

The test case: turn Yeats' *The Second Coming* into a first-person game. The agent went through planning, JSON scaffolding, HTML build, a 6-point Puppeteer validation loop, physics-failure triage, and visual-tone screenshot review. Final score: 11/11 validation checks passed, playable in browser, with the "beast" aggroing on the player. No clarifying questions were asked during the run.

The QC harness is the actual product

The agent didn't guess fixes — it recognized that the force logging only captures the first eight steps during the fall, then re-ran validation after each patch. On the trickiest check (structure fragments into tumbling bodies), it iterated from 9/11 → 10/11 → 11/11. The creator notes this verification step is much more so than designing the initial game and is where the money goes.

The real cost

The build consumed 13 million tokens and cost $32, chewing through 72% of his Cursor API quota in the second week of the month. His baseline estimate for a similar 3D project without loops was $5–$6 — loops roughly 2x the spend. He now explicitly recommends you plan out the cost before you use loop.

12-month outlook

Exactly 1 year from now, it's just going to be on a whole different level. Translation: this workflow is good enough to form as a habit today, not a future bet.