UI Evidence

Use this skill when the user wants an agent to:

• bootstrap ui-evidence into the current repo
• capture before and after UI screenshots
• capture the current UI into a reviewable snapshot bundle
• compare two UI states locally
• generate a review page a human can scan quickly
• persist a reusable config for future UI work
• compare the current checkout against another git ref

This skill is intentionally thin. The CLI is the engine. The skill is the agent-facing entrypoint. The package provides the repo-local executable that actually runs capture and review commands.

Bootstrap first

1. Detect the repo package manager from packageManager or lockfiles.
2. Check whether the repo can already run ui-evidence.
3. If the CLI is missing, install it from GitHub:

pnpm add -D github:0xBrewing/ui-evidence
npm install -D github:0xBrewing/ui-evidence
yarn add -D github:0xBrewing/ui-evidence
bun add -d github:0xBrewing/ui-evidence

1. Run the repo bootstrap step with the matching package runner:

pnpm exec ui-evidence install --agent both --config ./ui-evidence/config.yaml
npx ui-evidence install --agent both --config ./ui-evidence/config.yaml
yarn ui-evidence install --agent both --config ./ui-evidence/config.yaml
bunx ui-evidence install --agent both --config ./ui-evidence/config.yaml

That bootstrap should leave repo-local skill copies in .agents/skills/ui-evidence/ for Codex and .claude/skills/ui-evidence/ for Claude Code.

Default flow

1. Run ui-evidence discover.
2. Read the suggested config and unresolved list.
3. Ask only about unresolved values.
4. Create or patch ui-evidence/config.yaml.
5. Add ui-evidence/hooks/* only if deterministic state is needed.
6. Run ui-evidence doctor, then ui-evidence doctor --ready if route or wait-target confidence is still low.
7. Run ui-evidence run, ui-evidence snapshot, or capture/compare/report/review.
8. Summarize:

• review/index.html
• report.<lang>.md
• manifest.json
• pair and overview images, or current snapshot captures and overview images

Commands to prefer

ui-evidence discover --format json
ui-evidence init --interactive --config ./ui-evidence/config.yaml
ui-evidence doctor --config ./ui-evidence/config.yaml
ui-evidence doctor --config ./ui-evidence/config.yaml --ready --profile mobile-en
ui-evidence run --config ./ui-evidence/config.yaml --stage <stage-id>
ui-evidence run --config ./ui-evidence/config.yaml --stage <stage-id> --before-ref main
ui-evidence run --config ./ui-evidence/config.yaml --stage <stage-id> --after-attach http://127.0.0.1:3000 --resume
ui-evidence snapshot --config ./ui-evidence/config.yaml --scope <scope-id>
ui-evidence review --config ./ui-evidence/config.yaml --stage <stage-id>

ui-evidence review --stage <stage-id> reuses the latest snapshot current captures when a stage has no before/after comparison artifacts yet. If neither stage artifacts nor snapshot artifacts exist, it should fail instead of emitting an empty review.

What to inspect locally first

• package manager and dev scripts
• Playwright config and existing auth states
• harness routes or storybook routes
• stable data-testid targets
• existing screenshot folders
• whether the repo can boot from another git ref

Rules

• Prefer this skill as the install entrypoint when it is available through skills add.
• Use the package-manager-native runner for every ui-evidence command.
• Keep stage definitions stable and additive.
• Prefer existing harness routes over fragile live flows.
• Prefer testId waits over loose selectors.
• Treat review/index.html as the default human-facing artifact.
• Keep hooks small and deterministic.
• Do not re-implement browser automation in prompt text.