Claude Code vs Codex: Which Builds a Better App From One Prompt?
Codex and Claude Code built the same Typeform clone from an identical prompt. Here's how their cost, speed, and output quality compared.
What happens when Codex and Claude Code get the exact same prompt?
A developer ran an identical build prompt through both coding agents, asking each to create a production ready Typeform alternative using a research, build, and verify workflow. The two tools produced apps that differed sharply in both process and result. One agent leaned toward visual polish. The other leaned toward working functionality. Neither output was flawless, and the gap between them says a lot about where each tool currently stands as an autonomous app builder.
TL;DR
- Identical prompts produced very different builds: the same slash-command goal prompt, run through Codex and Claude Code separately, resulted in two apps with different names, different UI approaches, and different bug profiles.
- Codex’s build looked more polished on the surface, with a stronger landing page, a hero image, and visual details that read as more professionally designed at first glance.
- Claude Code’s build was less visually refined but far more usable, with a form-building flow that actually resembled Typeform and worked the way a user would expect.
- Both outputs shipped with real bugs, including broken image previews, unresponsive theme controls, mislabeled question numbers, and navigation dead ends inside settings menus.
- The build times and costs diverged enormously, with one run taking a matter of hours and the other stretching across days, and one costing roughly three times more than the other.
- The creator expected Codex to win based on prior experience, but Claude Code’s app was judged the clear winner on functionality and usability despite looking rougher.
- Prompt design mattered as much as model choice, since the creator noted that adding a dedicated planning phase between research and build likely would have improved both results.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
How was the test set up?
The comparison used a single prompt, delivered as a slash-command goal, given verbatim to both Codex and Claude Code. The prompt asked each agent to build a production ready, originally branded Typeform alternative, and to orchestrate the work across three phases: research, build, and verify. It explicitly instructed the agent not to stop at a first working prototype, but to keep testing, breaking, and fixing until the app reached a genuinely complete state.
Both agents worked from that same instruction set with no additional hand-holding. One run happened on a Mac while the creator was traveling, the other on a separate PC. Neither output was previewed or clicked through before the on-camera evaluation, so the reactions captured were the first real impressions of each build.
What did Codex’s app get right, and where did it fall short?
Codex’s build, named Real Form in the test, made the strongest first impression. The landing page had a real hero image, layered background depth, a pill-shaped UI element, and copy that felt considered. The signup flow was clean enough that an average visitor likely wouldn’t immediately clock it as AI-generated.
Once inside the actual form builder, though, the experience got busy fast. The editor exposed a lot of controls at once, variables and formatting options stacked on screen in a way that felt overwhelming rather than guided. Some of that complexity paid off on the backend: the tool supported a wide range of field types, including opinion scales, matrices, and file uploads, showing the agent had thought carefully about the kinds of data an admin might want to collect.
But usability suffered. An uploaded image showed a “preview unavailable” error. Confirmation dialogs popped up in visually broken positions. Theme customization controls, including color pickers, didn’t appear to actually change anything when clicked. And because the build treated the experience as a demo workspace without real authentication, some of the deeper functionality (webhooks, email notifications, live publishing) wasn’t fully wired up. The impression was a tool that nailed the look of a serious product but hadn’t stress-tested the actual workflow a form creator would use.
What did Claude Code’s app get right, and where did it fall short?
Claude Code’s build, named Fora, made a rough first impression. The landing page was described as visually plain, almost the opposite of Codex’s polished hero section. But once the creator moved past signup and into the actual product, the experience held up much better.
The core form builder felt intuitive and closer to Typeform’s real interface: a clear choice between conversational and stacked display modes, a progress bar, question numbering, keyboard hints, autosave, and partial-response capture. Multiple workspaces could be created and switched between. Field types included short text, email, phone, website, dropdown, picture choice, net promoter score, opinion scale, rating, and ranking, covering the same functional ground as Codex’s build but with a simpler, less cluttered editing experience.
One coffee. One working app.
You bring the idea. Remy manages the project.
It wasn’t bug-free. A settings button that should have surfaced a logout option didn’t work. Question numbering logic broke in places, showing “one of three” in one context and “one of one” in another. Clicking into certain fields sometimes required clicking away and back before the selection registered. The design and web-hook sections both trapped users in sub-screens with no easy way to navigate back. Despite that, the form the agent produced could actually be published, filled out, and reviewed: real responses showed up in a results dashboard with both summary and individual-submission views. That end-to-end loop, prompt to published form to captured data, worked, which is the core job a Typeform clone has to do.
Which tool actually won, and why does it matter?
Judged purely on functionality and the experience of building and publishing a working form, Claude Code’s app was the clear winner, even though it looked less professional out of the box. Codex’s app made a stronger first impression but broke down under actual use, with core interactions like theme changes and image previews failing outright.
This matters because the two failure modes represent different risks for anyone using AI coding agents to ship real products. A tool that produces polished but non-functional UI can fool a first-glance review, whether that’s a stakeholder demo or a quick QA pass, and only reveal its problems once real users start clicking around. A tool that produces plainer but more functional UI is easier to visually improve later, since design polish is a more contained problem than fixing broken interaction logic.
The creator also noted an important caveat: identical prompts don’t guarantee identical process quality. One run took a matter of hours, the other stretched across multiple days, and their costs differed by roughly three times, according to the comparison. Neither agent was given a dedicated planning phase between research and build, and the creator suspected that gap, more than any inherent model limitation, explains a meaningful share of the bugs in both outputs. That’s a reminder that prompt structure and phasing can shape results as much as the underlying model choice.
Is one of these tools simply better than the other?
Not in any absolute sense, based on this single test. What the comparison does show is a consistent pattern worth watching for: one agent optimizing toward visual presentation, the other toward functional completeness, when given the same open-ended instructions. Neither fully delivered on the “production ready, no bugs” bar set in the prompt. Both needed real human QA to catch issues like non-responsive buttons, broken navigation, and mislabeled fields, exactly the kind of testing the prompt asked the agents to do themselves.
The practical takeaway for developers is to treat either tool’s first output as a strong draft, not a finished product, and to budget time for the kind of manual click-through testing that surfaced every bug in this comparison.
Frequently Asked Questions
What app did Codex and Claude Code each build in this comparison?
Both were given the same prompt to build a production ready, original Typeform alternative. Codex produced an app called Real Form. Claude Code produced an app called Fora.
Which agent produced a better looking app?
Codex’s build had stronger visual polish, including a hero image and more considered landing page design, and looked more like a professionally designed product at first glance.
Which agent produced a more functional app?
Claude Code’s build was judged more functional. Its form editor was less cluttered, more closely resembled Typeform’s actual workflow, and successfully supported publishing a form and capturing real responses.
Did either app work perfectly out of the box?
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
No. Both had real bugs, including broken theme controls and image previews in Codex’s build, and navigation dead ends and numbering errors in Claude Code’s build.
Why did the two builds turn out so differently from the same prompt?
Part of the gap likely comes from how each agent interpreted an open-ended, multi-phase prompt without a dedicated planning step. The creator noted that adding a planning phase between research and build would likely improve results from either tool.