Evidence, not rankings

What Do AI Coding Agents Actually Build?

We give agents the same brief, then publish exactly what they built, screenshots, checks, and failures included.

9 experiments · Claude Code, Codex, MiniMax tested · every prompt public

Experiment

We gave Claude and Codex the same raw evidence. Which made it easier to understand?

For this experiment, we asked Claude and Codex to turn the same frozen evidence from an earlier test into a case-study page. We then compared how clearly each one communicated the result.

Full-page desktop capture of Claude Code’s evidence-to-case-study result, with the source question, two Relay build screenshots, detailed differences, checks, mobile evidence, interpretation, and raw evidence links. — Claude CodeCase-study pageOpen full capture

Full-page desktop capture of Codex’s evidence-to-case-study result, with a large editorial headline, scoped result, two prominent Relay build screenshots, meaningful differences, checks, takeaway, method, and evidence links. — CodexCase-study pageOpen full capture

Codex made the source experiment easier to understand in this run: its shorter page used a stronger headline-to-question-to-large-screenshot sequence and a more concise scoped takeaway, while Claude exposed more raw traceability but made the reader work through a longer, denser comparison.

Read the full experiment →

Experiment

Claude Code vs Codex

We gave Claude Code and Codex the exact same SaaS landing-page brief from the same starting point. Here is what each one built.

Full-page desktop capture of the Relay SaaS landing page Claude Code produced: navigation, hero headline "Run your team's ops on autopilot", social proof, feature sections, ROI calculator, pricing, FAQ, and footer. — Claude CodeRelay landing pageOpen full capture

Full-page desktop capture of the Relay SaaS landing page Codex produced: navigation, hero headline "Less busywork. More momentum.", social proof, four feature sections, ROI calculator, four pricing plans, FAQ, and footer. — CodexRelay landing pageOpen full capture

When we checked both builds against the same 20 requirements, Codex passed all 20 and Claude Code passed 18. We did not score which page looks better; that part is still yours to decide.

Read the full experiment →

Experiment

Can Claude Code build a Next.js app from one prompt?

One prompt, no follow-up questions. This is Signalboard, the dashboard that came back.

Desktop capture of the Signalboard dashboard Claude Code produced without CLAUDE.md: dark sidebar, four summary cards, a signals-by-type bar chart, recent activity list, and a status breakdown. — Claude CodeSignalboard, one promptOpen full capture

It built, typechecked, and rendered cleanly on desktop and mobile. On the dashboard route we checked, we could not verify the record-management features the prompt asked for.

Read the full experiment →

Experiment

Does CLAUDE.md actually help?

Same task, same starting commit, same model; the only declared difference was a root CLAUDE.md file.

Desktop capture of the Signalboard dashboard Claude Code produced with a root CLAUDE.md: light sidebar, four summary cards, a signals-by-type bar chart, and a recent signals list of tagged cards. — Claude CodeWith CLAUDE.mdOpen full capture

The two builds look different. On the checks we ran, neither version performed better than the other.

Read the full experiment →

Experiment

Claude Code Plan Mode vs Direct: Same Result, 57% More Time

Both Claude Sonnet 4.6 builds passed every functional check and tied on blinded UX review; native Plan Mode took 79.587 seconds longer in this Project Board test.

Desktop screenshot of the Direct arm's responsive Project Board with To Do, In Progress, and Done columns — Direct (desktop)Claude Sonnet 4.6 · 10/10 checks · 16/16 blinded UXOpen full capture

Desktop screenshot of the native Plan Mode arm's responsive Project Board with three task columns — Native Plan Mode (desktop)Claude Sonnet 4.6 · 10/10 checks · 16/16 blinded UXOpen full capture

In this single controlled Project Board test using Claude Sonnet 4.6, Claude Code's native Plan Mode produced the same measured correctness and UX outcome as Direct implementation, while taking about 57% longer and using substantially more of the exposed token fields.

Read the full experiment →

Experiment

Claude Code vs Codex: Front-End Design Test

Codex wins with medium confidence on this one redesign: stronger overall composition and a more distinctive desktop experience, with Claude clearly better on mobile usability and product-story fidelity.

Desktop screenshot of the frozen Trace baseline, a competent but visually plain AI code-review product page with a text-only hero, basic benefit cards, review rows, testimonial, and final CTA. — Baseline (desktop)Frozen starting page · 1440px viewportOpen full capture

Desktop screenshot of Claude Code's Trace redesign, with a restrained blue hero, realistic inline code-review mockup, icon-led benefit cards, review list, testimonial card, and dark blue CTA. — Claude Code (desktop)Claude Opus 5 · HighOpen full capture

In this single controlled redesign, Codex produced the stronger overall visual composition and more distinctive desktop experience, while Claude was more restrained, more faithful to the product story, and clearly better on mobile usability. Codex's win comes with two meaningful flaws: it invented a 94/100 product score and removed mobile section navigation.

Read the full experiment →

Experiment

Does a Frontend Skill Actually Make Codex Better?

Both builds passed 25/25 deterministic checks. The frontend skill changed Codex's architecture substantially, took about 29% longer, and did not produce the build a non-blinded human reviewer preferred.

Desktop screenshot of Codex's Relayboard Upgrade Team Plan flow built without the frontend skill, in a single largely monolithic 263-line page component. — Codex (no skill)Baseline arm · desktopOpen full capture

Desktop screenshot of Codex's Relayboard Upgrade Team Plan flow built with the frontend implementation skill applied, decomposed into separate step components, typed plan data, and a shared currency-formatting helper. — Codex (with frontend skill)Skill-treatment arm · desktopOpen full capture

On this one test, the frontend implementation skill did not produce a measurably more correct build: both arms passed all 25 automated checks. It did substantially change how Codex organized the code, moving it from one large page component toward the decomposed, typed, helper-based structure the skill prescribes. That architectural change cost about 29% more runtime and did not yield the build a human reviewer preferred in this non-blinded look. This is one paired run on one task with one skill and one Codex configuration; it does not show that skills generally help or hurt Codex, and it does not establish that the skill made Codex worse overall, only that it changed the tradeoff on this run.

Read the full experiment →

How We Test

For each experiment, we give the agents the same task from the same starting point. Then we check what they built against the same requirements, inspect the actual outputs, and publish the results, including what failed. We keep the prompt, setup, screenshots, and checks so you can inspect the evidence yourself.