| name | site-improvement-backlog |
| description | Turn loose ideas about this site's UX, SEO, performance, or maintainability into a prioritized, executable backlog. Use when the user asks to audit the site, propose improvements, or turn scattered feedback into actionable tasks — not for a single isolated bug fix. |
| metadata | {"ported-from":"ai/skills/site-improvement-backlog.SKILL.md"} |
Site Improvement Backlog
Adapted from this repo's ai/ context kit for direct use in Claude Code. Read ai/mcp/source-of-truth-policy.md before validating any recommendation that depends on external/current information (library versions, Azure guidance) rather than repo state alone.
When to use
- Evaluating the site's UX, performance, SEO, or reliability and proposing prioritized work
- Turning scattered ideas/feedback into an executable backlog
Require a specific focus area — refuse to produce a backlog for an unscoped "audit everything" without first narrowing scope (e.g. "SEO on article pages", "bundle size", "test coverage").
Priority levels
- P0: functional breakage, regression, deploy-blocking
- P1: SEO, accessibility, core performance
- P2: incremental UX/content improvements
- P3: low-risk visual/technical polish
Dimensions to check
- Content: clarity, consistency, technical currency — see
publishing-content skill for the Article model itself
- SEO: metadata completeness, structure, discoverability — see
seo-structured-data skill
- Performance: bundle size, rendering, image weight — see
performance-budget skill
- DevEx: code conventions, test coverage, CI pipelines — see
testing-frontend and deploying-swa-terraform skills
Reuse those skills' concrete facts (e.g. known i18n drift, missing article images, disabled E2E in CI) as backlog input instead of re-deriving them from scratch.
Output format
A prioritized list with: id, priority (P0-P3), title, impact, effort (low/medium/high), risk, and a one-line technical justification per item. Any recommendation with an external dependency (library version, Azure behavior) must be checked against current documentation before being listed, not assumed from training data.