Scenario: Greenfield project scaffolding
Compiled by struct.dev
No consensus, and we're not going to manufacture one. The evidence splits cleanly along what people mean by scaffolding: those who want a complete, opinionated skeleton in one shot report better results with Codex; those who want to shape architecture interactively from the first file prefer Cursor and consider Codex's output a rewrite-in-waiting. Read the points, decide which camp you're in, and know that both camps include people whose judgment we trust.
Confidence is labeled honestly: disputed — read the thread before choosing.
239 practitioners weighed in for the “Greenfield project scaffolding” scenario specifically
Strong evidence gets promoted into the record above.
The vacuous test point should be a site-wide banner honestly. I now have a standing rule: after any scaffold, run the suite with the implementation deleted. If tests still pass, they're not tests. Catches garbage from every agent, not just these two.
Mutation testing for the impatient. I do the same but with git stash on src/ — same idea, faster. Both agents fail this check on roughly a third of generated test files in my logs.
React Native data point: neither wins because both scaffold RN like it's 2023. Codex reached for a navigation lib version incompatible with the RN version it had itself pinned ten minutes earlier. Cursor at least let me catch it live. For mobile, camp Cursor by forfeit.
I scaffolded the same brief 5x with each for my suite. Codex variance: low, quality median: good. Cursor variance: huge, quality ceiling: highest single result of the ten. Pick based on whether you want the median or a shot at the ceiling. This comparison is DISPUTED because both camps are correct about the distribution.
Everyone arguing skeleton quality is missing that briefs beat both. My scaffold prompt is 300 lines of AGENTS.md at this point — stack pins, directory layout, the works. With that, Codex and Cursor converge to nearly the same output. The disagreement here is mostly measuring unprompted defaults, which is a choice you don't have to make.
Sign in to join the discussion, vote, and verify fixes.
Related records
Firmly camp Codex. We're 4 engineers; a working skeleton in 18 minutes that we reshape over the next week beats 70 minutes of me babysitting Cursor to get a skeleton I merely like more. Architecture at day zero is a luxury opinion — day-zero momentum is what kills startups.
If we're alive in 14 months to pay your rate, the skeleton did its job. But point taken.
And I make a living off teams who took the 18-minute skeleton and built 14 months on top of it. The fetching-in-components pattern Codex loves becomes a $40k consulting engagement at scale. Momentum in the wrong direction is still velocity, sure.