What agent skills do to a web page

Method

The model and effort are fixed in the run scripts. The table shows what the run.json files record.

One sample per cell, no picking

Each cell is generated once, in one pass, with no follow-up turns. The result is shown as it came out, failed runs included. If a cell is rerun, the earlier attempt is kept in runs/<prompt>/_attempts/.

Same prompt in all four cells

The prompts use neutral wording with no style adjectives. Skill cells add one closing sentence that names the skill: Use the <name> skill. for Claude Code and Use the $<name> skill. for Codex.

Isolation

Every run starts in a fresh throwaway directory outside the repository, with an empty HOME, its own config home (CLAUDE_CONFIG_DIR or CODEX_HOME) and an environment built from scratch. No user settings, plugins, user-level skills, MCP servers or memories are loaded. The skill under test is copied only into that throwaway project's skill folder (.claude/skills/ or .agents/skills/). Skills bundled with each CLI stay available and count as native. The evidence comes from each agent's own logs and is stored in run.json under isolation.

The skill has to fire

For skill cells, the logs must show that the skill was loaded (skill_triggered). If it was not, the cell is marked Failed instead of being shown as a skill result.

Screenshots

Full-page captures at 1440 and 390 CSS px wide, taken after the network is idle and fonts have loaded, with animations disabled (playwright 1.64.1, playwright-chromium 156.0.8078.4). Thumbnails show the top of the page. “Open page” loads the original index.html in a sandboxed iframe (sandbox="allow-scripts", no same-origin access).

Accessibility scan

Each page is also checked with an automated accessibility scanner, and the number of violations by severity is shown. Automated scans catch only part of accessibility problems (they cannot judge, for example, whether alt text or focus order make sense), so treat the counts as a pointer, not a verdict.

Scroll recordings

Some cells also have a short recording of the page being scrolled at desktop width, because motion does not show in screenshots. It loads only when you press play on the cell page.

Run dates

2026-10-25 – 2026-10-26 (UTC), 46 cells.

Votes

Each result page has a 👍 vote through giscus, a third-party widget that loads only on result pages, not on the rest of the gallery. Voting needs a GitHub sign-in. Votes are stored as reactions in this repository's GitHub Discussions, one discussion per cell, shared by all languages. The counts on the comparison pages are refreshed every few hours. They show what visitors prefer; they are not an evaluation.

Known limitations

• Native means out of the box, not zero skills. Every cell, native ones included, can use the skills that ship with its agent: the 19 skills bundled with Claude Code, and Codex's system skills (imagegen, openai-docs, skill-creator, skill-installer). In the dashboard prompt, Claude Code's bundled dataviz skill fired on its own twice, once in claude__native and once in claude__frontend-design; run.json records this under builtin_skills_used. This is on purpose: the comparison is about what a reader actually gets after installing the agent.
• Isolation means “nothing loaded”, not “nothing readable”. The agents can still read the rest of the disk: Codex's stock workspace-write sandbox can read /, and Claude Code with bypassPermissions has no sandbox. In the v0 runs, agents looked for AGENTS.md in parent directories and one used Playwright from the operator's npm cache. Stronger isolation (a separate user account or a container) is future work.
• Fixed sidebars in screenshots. Full-page captures draw position: fixed or sticky elements sized 100vh at the first screen's height only, so a dashboard sidebar can look as if it stops at 900 px. This comes from the capture, not from the page.
• One sample is not a benchmark. Each cell is a single generation. Treat the gallery as examples of what a skill tends to do, not as a score.