Diagnose Vercel CDN cache hit rates, stale content, revalidation behavior, and ISR costs.
npx skills add https://github.com/openai/plugins --skill cdn-caching
Please help me install this Agent Skill. Skill: cdn-caching Source: https://github.com/openai/plugins/tree/main/plugins/vercel/skills/cdn-caching Command: npx skills add https://github.com/openai/plugins --skill cdn-caching
Debugging why Vercel-served pages are stale, uncached, or expensive is hard because caching spans multiple layers (CDN, ISR, PPR, function invocations) with opaque per-request behavior. This Skill provides expert guidance and concrete CLI workflows to measure cache hit rates, explain per-request cache outcomes, and control ISR revalidation costs.
vercel metrics, excluding BYPASS noise, and drill into which routes and paths miss most.x-vercel-cache statuses and the nine cacheReason values (cold, collapsed, draft_mode, stale_tag, etc.) to explain exactly why a request missed or went stale.cache_tags, discover a broad blogPost tag invalidating thousands of pages, and replace it with granular blogPost:<id> invalidation.Ask the agent to analyze why my Vercel site's cache hit rate dropped and which routes are driving ISR write costs.
High-intent search queries and answers about installing and using this skill.
Run `vercel metrics vercel.request.count` grouped by `cache_result` for your team and project, filtered to production. Treat HIT, STALE, and PRERENDER as cache-served and exclude BYPASS and unset values so non-cacheable traffic does not distort the ratio.
Stale responses come from stale-while-revalidate behavior after time-based intervals or tag invalidation. Check the `cacheReason` field: `stale_time` means a scheduled refresh, `stale_tag` means an invalidation call, and `stale_error` means regeneration failed and the last-good copy is served.
Invalidate (revalidateTag, invalidateByTag) uses stale-while-revalidate, serving the old copy while regenerating in the background. Dangerously-delete (updateTag, dangerouslyDeleteByTag) hard-removes the entry, so the next request blocks in the foreground to regenerate.
No. A PPR route has dynamic holes by definition, so even a cached shell hit still invokes the function to render the dynamic parts. Only a route with no holes serves as a pure prerender HIT without a function call.
Focus on write_units, which bill per 8 KB on every revalidation. Prefer on-demand tag-based revalidation over short time-based intervals, scope tags to specific IDs like `product-<id>` instead of broad tags, and lengthen revalidate intervals where regeneration outpaces reads.
The largest legitimate BYPASS sources are Draft Mode, prerender-bypass cookies, and SEO crawlers, all of which intentionally skip caching. Group BYPASS requests by bot_category, user_agent, and request_method to confirm what remains before tuning headers.
vercel-labsAudits deployed Vercel projects for cost and performance using observability metrics and deterministic gates.
hopeovertureAnalyze Next.js routes to determine rendering strategies and revalidation intervals.
ChristineThamConfigure Vercel Firewall rules for DDoS mitigation and bot management.
cosmicstack-labsDesign caching strategies for backend systems using Redis and invalidation methods.