Webperf Skills & Diagnostics
Digitising an audit methodology into skills for AI agents. A 4-step pipeline (collection, causal diagnosis, scoring against 112 criteria, prioritisation) grounded in measured data.
- Status
- prototype
- Year
- 2025
- Stack
- Claude Code, Cursor, Skills, Web performance
Why this project
A proven web performance audit methodology stays, by default, in the head of the person who practises it. Digitising it into “skills” an AI agent can use — Claude Code, Cursor — makes it reproducible across a large number of sites, without reinventing the diagnosis on every engagement.
Use cases
- Automated collection of performance data on a target site
- Causal diagnosis: linking a measured symptom (high LCP, unstable CLS) to its likely cause
- Scoring against 112 criteria, grounded in actually measured data rather than a declarative checklist
- Prioritising fixes by estimated impact
How it works
The pipeline runs in 4 steps: metric collection, agent-assisted causal diagnosis, scoring against the methodology’s 112 criteria, then action prioritisation. Each step is encapsulated in its own skill that the agent loads on demand, rather than one monolithic prompt — keeping every step independently testable and fixable.
What I took away
Digitising an audit methodology into structured skills, rather than a generic prompt, forces the diagnostic reasoning to be made explicit step by step — which benefits the agent as much as the clarity of the methodology itself.