What it does. Use for Vercel cost and performance optimization on deployed projects, especially Next.js, SvelteKit, Nuxt, and limited Astro apps. Collect Vercel metrics, usage, project config, and code scan results first; investigate only metric-backed candidates; produce ranked recommendations grounded in verified files and version-aware Vercel/framework docs. Trigger for Vercel bill reduction, slow or expensive routes, caching opportunities, Function Invocations, Build Minutes, Fast Data Transfer, Core Web Vitals, Bot Management, Fluid compute, or cost breakdown requests. Part of vercel-labs/agent-skills (Vercel official skills) (vercel-labs/agent-skills).

Upstream	vercel-labs/agent-skills
Skill file	skills/vercel-optimize/SKILL.md
License	MIT (stated in the README; no LICENSE file)
Author	Vercel Labs
Fetched	2026-09-10

Install

• npx skills add vercel-labs/agent-skills --skill vercel-optimize, or copy the skill folder into ~/.claude/skills/vercel-optimize/.
• Raw file: curl -sL https://raw.githubusercontent.com/vercel-labs/agent-skills/HEAD/skills/vercel-optimize/SKILL.md

SKILL.md (verbatim)

name: vercel-optimize
description: "Use for Vercel cost and performance optimization on deployed projects, especially Next.js, SvelteKit, Nuxt, and limited Astro apps. Collect Vercel metrics, usage, project config, and code scan results first; investigate only metric-backed candidates; produce ranked recommendations grounded in verified files and version-aware Vercel/framework docs. Trigger for Vercel bill reduction, slow or expensive routes, caching opportunities, Function Invocations, Build Minutes, Fast Data Transfer, Core Web Vitals, Bot Management, Fluid compute, or cost breakdown requests."
metadata:
  version: "1.2.0"

Vercel Optimize

Run an observability-first Vercel optimization audit. Do not inspect source files until signals.json exists and a deterministic gate points to a route, file, or project setting.

Core doctrine: read references/doctrine.md if any rule is unclear.

• Metrics first. Recommendations start from Vercel production signals, not repo-wide grep.
• Deterministic gates. scripts/gate-investigations.mjs decides what deserves investigation.
• Candidate-bound scope. Read only files named by a candidate or a route-local import chain.
• Version-aware citations. Use only references/docs-library.json; invalid or version-mismatched citations are stripped.
• Customer copy. Read references/voice.md before writing report text or chat output.

Prerequisites

• Vercel CLI v53+ with vercel metrics, vercel usage, vercel contract, and vercel api.
• Authenticated CLI session: vercel login.
• Linked app directory: vercel link. VERCEL_PROJECT_ID can help resolve project config, but vercel metrics still requires directory linkage. The link or environment must include the intended project org/team/user scope so the collector can resolve a CLI-safe --scope and keep vercel metrics, vercel usage, and vercel contract on the same account.
• Node.js 20+.
• Observability Plus for route-level metric-backed recommendations.

Never put auth tokens in shell commands. Do not type VERCEL_TOKEN=..., --token ..., or Authorization: Bearer ... into commands that may be echoed in chat.

Framework Support

The preflight reads package.json and sets expectations before metric fan-out.

Framework	Status	Notes
Next.js App Router	supported	strongest route mapping, scanners, playbooks, citations
Next.js Pages Router	supported	scoped to Pages Router idioms when detected
SvelteKit	supported	route mapping for src/routes files and SvelteKit scanner
Nuxt	supported	route mapping plus generic/platform checks; fewer framework-specific recs
Astro	limited	route mapping plus generic checks; fewer framework-specific recs
Hono / Remix / unknown	blocked by default	continue only if the user accepts a limited platform/code-only audit

If unsupported, stop and ask before scanning or gating:

This project uses <framework>. Vercel Optimize supports metric-backed code recommendations for Next.js, SvelteKit, and Nuxt. Astro support is limited. For <framework>, I can still run a limited platform/scanner audit, but route-level Vercel metrics may not map back to source files.

Do you want me to continue with the limited audit, or stop here?

If the user continues, rerun collection with --continue-unsupported-framework.

Run Directory

Use a fresh run directory for every audit. Do not reuse briefs, sub-agent outputs, or reports across runs.

RUN_DIR="$(mktemp -d -t vercel-optimize-XXXXXX)"

Pipeline

1. Collect, scan, and merge signals

Run from the linked app directory or pass --cwd where a script supports it. Keep stdout JSON separate from stderr logs. Do not combine streams.

node scripts/collect-signals.mjs [projectId] > "$RUN_DIR/vercel-signals.json" 2> "$RUN_DIR/collect.stderr"
node -e 'JSON.parse(require("fs").readFileSync(process.argv[1], "utf8"))' "$RUN_DIR/vercel-signals.json"

node scripts/scan-codebase.mjs <repo-root> > "$RUN_DIR/codebase.json"
node scripts/merge-signals.mjs "$RUN_DIR/vercel-signals.json" "$RUN_DIR/codebase.json" --out "$RUN_DIR/signals.json"

Collection details, schemas, metric IDs, and degradation behavior live in references/data-collection.md. The metric registry is lib/queries.mjs; keep all queries on the shared 14-day window.

collect-signals.mjs resolves the linked project owner to commandScope.cliScope and verifies that the resolved account can read the resolved project before it checks Observability Plus. Downstream scripts reuse that scope for every Vercel CLI command that accepts --scope. Do not run vercel usage, vercel metrics, or vercel contract manually without the same scope; unscoped usage can report the user's personal organization while route metrics come from the team project.

If project or scope resolution is ambiguous, stop and ask the user which Vercel project and team/personal scope they want audited. Do not infer the intended scope from the current vercel whoami team, and do not proceed with metrics, usage, or contract collection until the link, an exact project match in .vercel/repo.json, or VERCEL_PROJECT_ID + VERCEL_ORG_ID identifies the intended account.

Use this prompt for PROJECT_SCOPE_UNRESOLVED, SCOPE_UNRESOLVED, or PROJECT_SCOPE_MISMATCH:

I can't safely identify the Vercel project and account for this audit yet.

Please confirm the Vercel project name or ID and the team slug/name, or tell me it's under your personal account. Once confirmed, I'll relink or rerun collection against that exact scope before checking metrics.

1.1 Stop on blockers

Check blockers before gating:

jq '{frameworkSupportBlocker, observabilityPlus, observabilityPlusUsable, observabilityPlusBlocker, observabilityPlusBlockerDetail}' "$RUN_DIR/signals.json"

Required actions:

• frameworkSupportBlocker === "unsupported_framework": use the unsupported-framework prompt above.
• PROJECT_SCOPE_UNRESOLVED, SCOPE_UNRESOLVED, or PROJECT_SCOPE_MISMATCH: stop and ask which Vercel project and team/personal scope the user wants audited. For team projects, rerun after vercel link --yes --project <project-name-or-id> --team <team-slug>; for personal projects, rerun after linking under the intended user account or after setting both VERCEL_PROJECT_ID and VERCEL_ORG_ID.
• observabilityPlusBlocker === null: continue.
• no_traffic: tell the user route metrics are sparse; continue only if they accept limited output.
• payment_required or no_oplus_probe: render references/observability-plus.md verbatim and ask.
• project_disabled: tell the user to enable Observability Plus for the project or accept a limited audit.
• daily_quota_exceeded: stop and tell the user the Observability query quota is exhausted; retry after the next UTC midnight reset, or ask whether to continue with a limited code-only audit.
• not_linked: link the app directory, then rerun Step 1. If app path and project are known:

vercel link --yes --project <project-name-or-id> --cwd <app-dir>
# add --team <team-id-or-slug> when known

• forbidden or project_not_found: fix auth/team scope. Do not pitch Observability Plus.
• all_failed_other: show the raw error code and ask whether to continue in limited code-only mode.

Do not silently fall back to code-only mode. If the user accepts a limited audit, rerun collection with:

node scripts/collect-signals.mjs [projectId] --continue-without-observability > "$RUN_DIR/vercel-signals.json" 2> "$RUN_DIR/collect.stderr"

Then scan and merge again.

2. Gate candidates

node scripts/gate-investigations.mjs "$RUN_DIR/signals.json" > "$RUN_DIR/gate.json"

Output shape:

• toLaunch: code-scope candidates to investigate.
• platform: project/account-scope recommendations.
• gated: skipped, covered, or disqualified candidates that must still appear in the report.
• budget: candidate budget and selection mode.

Default budget is 6 code-scope candidates with a diversity guardrail. To expand:

node scripts/gate-investigations.mjs "$RUN_DIR/signals.json" --max-candidates 12 > "$RUN_DIR/gate.json"
node scripts/gate-investigations.mjs "$RUN_DIR/signals.json" --max-candidates all > "$RUN_DIR/gate.json"

Generated candidate docs: references/candidates.md.

2.1 Ask about audit scope when needed

Before deep-dive, run:

node scripts/budget-summary.mjs "$RUN_DIR/gate.json" --format json > "$RUN_DIR/budget-summary.json"

If shouldAsk is false, continue.

If shouldAsk is true:

1. Print exactChatMessage.body exactly as returned. Do not summarize, truncate, reorder, or rewrite it.
2. Then ask questionText using questionPayload when the host supports structured questions.
3. If the user chooses a different number, rerun the gate with --max-candidates <choice>.

Never put the long preview inside the question field. The preview and the question are separate surfaces.

2.2 Deep-dive and reconcile

node scripts/deep-dive.mjs "$RUN_DIR/signals.json" "$RUN_DIR/gate.json" --cwd <project-dir> > "$RUN_DIR/investigation-evidence.json"

node scripts/reconcile-candidates.mjs "$RUN_DIR/investigation-evidence.json" \
  --gate "$RUN_DIR/gate.json" \
  --out "$RUN_DIR/reconciled-investigation.json"

--cwd must be the linked project directory so deep-dive.mjs can verify the same project link and reuse signals.json.commandScope.cliScope for any follow-up vercel metrics calls.

Reconciliation deterministically converts disproven candidates into observations before any source investigation:

• metric_mismatch
• error_storm
• deployment_regression
• scanner_only_no_metric

2.3 Generate briefs and investigate

List the work:

node scripts/prepare-investigation-brief.mjs "$RUN_DIR/signals.json" "$RUN_DIR/reconciled-investigation.json" --list > "$RUN_DIR/briefs-manifest.json"

Generate one brief for every entry in briefs-manifest.json.briefs. The group can be toLaunch or platform; do not generate only toLaunch briefs.

mkdir -p "$RUN_DIR/briefs" "$RUN_DIR/sub-agent-outputs"
node scripts/prepare-investigation-brief.mjs "$RUN_DIR/signals.json" "$RUN_DIR/reconciled-investigation.json" \
  --group <brief.group> --index <brief.index> --out "$RUN_DIR/briefs/<brief.group>-<brief.index>.md"

Use briefs-manifest.json.briefs[].label for visible worker names, for example Low cache-hit route on /docs/llm-digest/[...slug], not toLaunch-7.

Fan-out rule:

• 1-2 briefs: investigate inline.
• 3+ briefs: spawn one sub-agent per brief when the host supports it.
• Hosts without sub-agents: run inline serially.

Sub-agent contract:

• The brief is the whole prompt.
• Read only files listed in the brief, plus route-local imports when needed.
• Emit one JSON recommendation or one JSON no-change finding using references/recommendations.md.
• Do not cite URLs outside the provided citation subset.
• Do not recommend framework features unavailable in the detected version.

If a sub-agent reaches for repo-wide grep, the candidate is malformed; drop or abstain rather than widening scope.

2.4 Collect outputs

Save each raw investigation result in $RUN_DIR/sub-agent-outputs/, then collect:

node scripts/collect-sub-agent-outputs.mjs \
  --manifest "$RUN_DIR/briefs-manifest.json" \
  --out "$RUN_DIR/recommendations.json" \
  "$RUN_DIR/sub-agent-outputs/"

The collector extracts JSON, prepends pre-resolved records, enforces manifest order, and fails on missing, duplicate, unknown, or mismatched candidateRef values.

3. Verify recommendations

node scripts/verify-and-regen.mjs "$RUN_DIR/recommendations.json" \
  --signals "$RUN_DIR/signals.json" \
  --repo-root <project-dir> \
  --out "$RUN_DIR/verify.json"

This script extracts claims, verifies files/citations/version fit, grades quality, applies sanitizers, emits verifiedRecommendations, withheldRecommendations, renderableRecommendations, and creates regenPlan for failed or unsafe recommendations.

Recommendation schema, writing rules, sanitizer order, and grading rules: references/recommendations.md. Verification rules: references/verification.md.

For each regenPlan entry, rerun the same brief with a Previous attempt failed these checks section listing topFailures. Keep the regenerated output only if verification improves without gutting citations.

4. Render report and final message

node scripts/render-report.mjs "$RUN_DIR/verify.json" "$RUN_DIR/gate.json" "$RUN_DIR/signals.json" \
  --project <name> \
  --out "$RUN_DIR/report.md" \
  --message-out "$RUN_DIR/final-message.json"

Use --debug-out "$RUN_DIR/debug.json" only when developing the skill. Customer Markdown and chat output must not expose passRate, quality, sanitizer trails, raw sub-agent names, or other implementation fields.

After rendering, print final-message.json.body verbatim and stop. Do not add highlights, debug notes, raw counts, sub-agent summaries, or extra explanation. Render-time dedupe, platform caps, and hard-safety drops can change the customer-visible count, so never summarize from raw verify.json.

Report structure and impact framing: references/scoring.md.

Recommendation Rules

Every recommendation must:

• Trace to a launched candidate, platform candidate, pre-resolved observation, or verified traffic-independent scanner finding.
• Include observed metric evidence from signals.json or evidence.deepDive.
• Cite verified files with line numbers when code is involved.
• Include at least one allowed citation that applies to the detected framework/version.
• Use precise observed performance numbers.
• Use cost magnitude phrases only; never customer-facing $N savings.
• Do not recommend duration reductions for Vercel Workflow runtime endpoints (/.well-known/workflow/v1/*). These are generated orchestration routes for durable step/flow execution and should be hard-gated before investigation.
• Workflow recommendations must name the boundary being changed. Valid examples: enqueue durable work and return a run ID instead of awaiting completion, fix stream replay/closure/locks, or reduce verified excess Workflow Steps/Storage. Do not infer cost savings from Workflow endpoint wall-clock duration.
• For streaming, SSE, resumable chat, or other intentionally long-lived routes, do not frame wall-clock function duration as a problem by itself. Require evidence of avoidable pre-first-byte work, high active CPU, duplicate invocations, or post-response work that can move out of the user-visible path.
• Name a specific cache policy when recommending caching.
• Keep unsafe responses dynamic unless evidence proves they are safe to cache: auth-sensitive paths, errors, fallback responses, missing content, invalid requests, geolocation/device-varying output, and unversioned dynamic URLs.

Never recommend "verify X is on" for facts already present in signals.project, including Fluid compute status, memory tier, regions, in-function concurrency, and timeout.

Scanner Rules

Scanner findings are supplementary. Drop findings annotated COLD-PATH or NO-ROUTE-MAPPING unless the scanner declares metadata.trafficIndependent === true.

Traffic-independent examples: middleware matcher, source maps, React Compiler config, build settings. Route-local cache or data-fetch patterns need route-level traffic evidence.

Scanner docs: references/scanner-patterns.md.

Final Customer Terms

Use:

• recommendations ready
• observations from investigation
• investigated, no change recommended
• not investigated in this run

Avoid:

• sub-agent
• abstention
• passRate
• quality score
• gate
• LLM

Failure Copy

Use these messages without adding sales copy or process detail.

No traffic in the last 14 days:

This project has no meaningful traffic in the last 14 days, so route-level metrics are sparse. I can still check traffic-independent scanner findings and project settings, but I cannot rank route fixes until traffic accumulates.

Route-level metrics unavailable:

Use the verbatim choice template in references/observability-plus.md. Do not silently fall back to code-only mode; present the two-path choice: enable Observability Plus and rerun the metric-backed audit, or accept a limited code-only run.

Project is not linked:

This worktree is not linked to a Vercel project. Run vercel link --yes --project <project-name-or-id> --cwd <app-dir> and rerun the audit. If the team is known, add --team <team-id-or-slug>.

Most route-to-file mappings failed:

The route inventory matched fewer than half of the routes we saw in observability. This is common in monorepos with custom routing. I've surfaced what I can match; the rest appear in the "Not investigated in this run" section.

Other files in this skill

• AGENTS.md
• CONTRIBUTING.md
• README.md
• lib/auth-route.mjs
• lib/budget-summary.mjs
• lib/citations.mjs
• lib/cost-coverage.mjs
• lib/dedup-recs.mjs
• lib/deep-dive.mjs
• lib/display-labels.mjs
• lib/extract-claims.mjs
• lib/framework-support.mjs
• lib/gates/build-minutes-fanout.mjs
• lib/gates/cold-start.mjs
• lib/gates/contract.mjs
• lib/gates/cwv-poor.mjs
• lib/gates/external-api-slow.mjs
• lib/gates/hard-gates.mjs
• lib/gates/index.mjs
• lib/gates/isr-overrevalidation.mjs
• lib/gates/middleware-heavy.mjs
• lib/gates/observability-events-attribution.mjs
• lib/gates/platform-bot-protection.mjs
• lib/gates/platform-fluid-compute.mjs
• lib/gates/region-misconfig.mjs
• lib/gates/route-errors.mjs
• lib/gates/scanner-driven.mjs
• lib/gates/select-candidates.mjs
• lib/gates/slow-route.mjs
• lib/gates/types.d.ts
• lib/gates/uncached-route.mjs
• lib/gates/usage-spike-triage.mjs
• lib/grade-recommendation.mjs
• lib/impact-label.mjs
• lib/impact-magnitude.mjs
• lib/investigation-brief.mjs
• lib/observation-safety.mjs
• lib/project-facts.mjs
• lib/queries.mjs
• lib/reconcile-candidates.mjs
• lib/render-report.mjs
• lib/repo-root.mjs
• lib/route-normalize.mjs
• lib/sanitizers/bot-protection-certainty.mjs
• lib/sanitizers/cache-tag-invalidation-certainty.mjs
• lib/sanitizers/count-correct.mjs
• lib/sanitizers/function-duration-invocations.mjs
• lib/sanitizers/index.mjs
• lib/sanitizers/middleware-conflict.mjs
• lib/sanitizers/missing-citation.mjs
• lib/sanitizers/pre-release.mjs
• lib/sanitizers/rate-limit.mjs
• lib/sanitizers/rendering-mode-mislabel.mjs
• lib/sanitizers/undeclared-dep.mjs
• lib/sanitizers/vercel-directive-strip.mjs
• lib/sanitizers/window-units.mjs
• lib/scanners/cache-components-suspense-dedupe.mjs
• lib/scanners/edge-heavy-import.mjs
• lib/scanners/force-dynamic.mjs
• lib/scanners/headers-in-page.mjs
• ... and 95 more (see the folder)

AGENTS.md (verbatim)

vercel-optimize

Cross-agent entry point for the Vercel Optimize skill. The full procedure is in SKILL.md.

Use this skill when the user asks to optimize a Vercel project, reduce a Vercel bill, investigate slow or expensive routes, find caching opportunities, reduce function invocations, or produce a Vercel cost/performance report.

Do not use it for projects that are not deployed on Vercel, greenfield projects with no traffic, or general code review.

Requirements

• Node.js 20+
• Vercel CLI with vercel metrics, vercel usage, vercel contract, and vercel api support; v53+ is this skill's compatibility floor
• Authenticated Vercel CLI session
• Linked Vercel project directory (vercel link) for route metrics. VERCEL_PROJECT_ID can help resolve project config, but it does not replace directory linkage for vercel metrics. The project must resolve to a CLI-safe team or personal scope so vercel metrics, vercel usage, and vercel contract all run against the same account.
• Observability Plus for per-route metric analysis

Procedure

1. Read SKILL.md.
2. Collect Vercel signals before reading source files.
3. Gate candidates with deterministic scripts.
4. Investigate only files named by launched candidates.
5. Verify recommendations mechanically before rendering the report.

The hard rules are in references/doctrine.md: observability first, deterministic gates, candidate-bound scope, and version-aware citations.

Install

Preferred:

npx skills add vercel-labs/agent-skills --skill vercel-optimize

Manual project install:

mkdir -p .agents/skills
cp -R <agent-skills-repo>/skills/vercel-optimize .agents/skills/

Then add this to the project AGENTS.md:

When optimizing Vercel cost or performance, follow
`.agents/skills/vercel-optimize/SKILL.md` before proposing changes.
Collect Vercel metrics before reading source files.

CONTRIBUTING.md (verbatim)

Contributing to vercel-optimize

Keep changes small, metric-grounded, and fixture-tested. Runtime code lives in skills/vercel-optimize; tests and fixtures live in packages/vercel-optimize-tests so installed skills stay small.

Common changes

Change	Edit	Test
Gate	lib/gates/<id>.mjs, lib/gates/index.mjs	node --test packages/vercel-optimize-tests/test/*gate*.test.mjs
Scanner	lib/scanners/<id>.mjs, lib/scanners/index.mjs	Scanner-specific test in packages/vercel-optimize-tests/test/
Citation	references/docs-library.json	node skills/vercel-optimize/scripts/check-citations.mjs
Support topic	references/support-topics/<id>.md	node --test packages/vercel-optimize-tests/test/support-topics.test.mjs
Playbook	references/playbooks/<profile>.md and selection matrix in references/scoring.md	node --test packages/vercel-optimize-tests/test/support-topics.test.mjs packages/vercel-optimize-tests/test/investigation-brief.test.mjs
Renderer or verifier	lib/render-report.mjs, lib/verify-claim.mjs, or related module	Focused test plus full test suite

Generated docs:

node skills/vercel-optimize/scripts/build-docs.mjs
node skills/vercel-optimize/scripts/check-docs-fresh.mjs

Full test loop:

node --test packages/vercel-optimize-tests/test/*.test.mjs
node skills/vercel-optimize/scripts/check-docs-fresh.mjs
node skills/vercel-optimize/scripts/check-citations.mjs

Rules

• No runtime dependencies. Scripts use Node.js 20+ built-ins and the Vercel CLI.
• No recommendation without a Vercel metric signal, code evidence when code changes are proposed, and an allow-listed citation.
• No invented URLs, exact savings projections, or version-mismatched framework APIs.
• No internal repo paths, service names, customer names, or captured private output in fixtures.
• Keep generated report copy customer-facing. Put debug details behind --debug-out.

Output contracts

Every JSON-emitting script must be deterministic: stable key order, stable sort order, 2-space indentation, trailing newline. If a consumed schema changes, update the schema version and the fixture tests in the same PR.

README.md (verbatim)

vercel-optimize

Optimize cost and performance for supported projects on Vercel.

This skill uses Vercel metrics to find high-impact improvements in your app. Every recommendation is backed by observed data, scoped code evidence, and version-aware docs.

Install

Install just this skill:

npx skills add vercel-labs/agent-skills --skill vercel-optimize

Manual install: copy skills/vercel-optimize into .agents/skills/vercel-optimize and reference SKILL.md from your project AGENTS.md.

Requirements

• Node.js 20+
• Vercel CLI with vercel metrics, vercel usage, vercel contract, and vercel api support (npm i -g vercel@latest). The skill enforces v53+ as its compatibility floor.
• Authenticated Vercel CLI session (vercel login)
• Linked Vercel project directory (vercel link) for route metrics. VERCEL_PROJECT_ID can resolve project config, but it does not replace directory linkage for vercel metrics. The project must resolve to a CLI-safe team or personal scope so vercel metrics, vercel usage, and vercel contract all run against the same account.
• Observability Plus for metric-backed route ranking
• Code-backed recommendation coverage is strongest for Next.js and SvelteKit, supported for Nuxt route mapping with generic checks, and limited for Astro. Hono, Remix, and unknown frameworks pause up front.

If route-level metrics are unavailable, the skill pauses before scanner-only mode. Scanner-only can catch traffic-independent code issues, but it cannot rank hot routes or prove cost impact.

Use

From the Vercel project directory, ask your coding agent:

optimize this Vercel project

The agent should collect metrics first. If it starts by reading source files or guessing from vercel.json, the skill was not loaded correctly.

Roadmap

Attribute	Status
Route-level Vercel Function invocations, duration, TTFB, and cold starts	Supported
Vercel Function CPU, memory, and GB-hours	Supported
Request volume, cache hit rate, HTTP status, and method distribution	Supported
Fast Data Transfer and bot traffic patterns	Supported
ISR reads, writes, and over-revalidation	Supported
Routing Middleware volume and duration	Supported
External API latency, volume, and transfer bytes	Supported
Core Web Vitals from Speed Insights	Supported
Image Optimization usage, source hosts, and source bytes	Supported
Build Minutes fan-out	Supported
Usage spikes by billing service	Supported
Bot Protection and BotID configuration	Supported
Fluid Compute configuration and compute signals	Supported
Region pinning and project configuration mismatches	Supported
Observability Events cost attribution	Supported
Route-to-file recommendations for Next.js and SvelteKit	Supported
Nuxt route mapping with generic/platform checks	Supported
Generic route mapping and platform checks for Astro	Supported
Hono route-to-file mapping	Planned
Remix route-to-file mapping	Planned
AI Gateway usage and cost optimization	Planned
Sandbox usage and cost optimization	Planned
Blob, Edge Config, Runtime Cache, Workflows, Queues, Flags, and Microfrontends billing dimensions	Planned

What You Get

• Ranked recommendations tied to observed Vercel metrics
• Specific route and file references when source changes are justified
• Before/after code for ready recommendations
• Citations from a curated, version-aware documentation allow-list
• Held-back findings when evidence is real but not strong enough for a recommendation
• A concise final message plus a full Markdown report

Trust Model

• Metrics come first. Code investigation starts only after signals are collected.
• Gates are deterministic JavaScript thresholds. No LLM decides whether a metric qualifies.
• Citations are allow-listed. Unknown URLs and version-mismatched framework docs are stripped.
• Project config contradictions are rejected. For example, the verifier blocks "enable Fluid Compute" when Fluid Compute is already on.
• Cost impact uses magnitude framing, not invented exact savings.

Contributing

See CONTRIBUTING.md. New gates, scanners, playbooks, citations, and sanitizers need fixture coverage in packages/vercel-optimize-tests.

License

MIT

references/candidates.md (verbatim)

<!-- THIS FILE IS GENERATED by scripts/build-docs.mjs. Do not edit by hand. --> <!-- To change scanner descriptions, edit lib/scanners/*.mjs metadata exports. --> <!-- To change gate thresholds, edit lib/gates/*.mjs metadata exports. -->

Candidate gates

The deterministic threshold expressions that turn observability signals into investigation candidates. Pure JS, no LLM. Thresholds live in lib/gates/*.mjs.

Total gates: 15. Budget cap: MAX_CODE_CANDIDATES = 6. Gate version: 1.8.0.

Gates

build_minutes_fanout

• Threshold: Build Minutes share > 0.15 OR turbo-force-bypass finding present
• Billing dimension: build
• Scope: account
• Source citation: vercel-optimize gate threshold

Build Minutes line dominates the bill or Turborepo cache is bypassed. On monorepos, unchanged work should be skipped through Vercel skip-unaffected behavior, a verified Ignored Build Step, and a complete Turbo cache contract.

---

cold_start

• Threshold: coldPct > 0.4 AND total >= 1000
• Billing dimension: function-duration
• Scope: route
• Source citation: vercel-optimize gate threshold

Routes where > 40% of invocations are cold-start, at meaningful traffic (>=1,000 total invocations in window). Cold starts add 200-800ms per request and break the perceived latency budget on cache-miss paths. The 40% threshold is where cold-rate becomes a real signal vs Poisson noise on serverless. Sourced from vercel.function_invocation.count grouped by function_start_type.

---

cwv_poor

• Threshold: LCP p75>2500 OR INP p75>200 OR CLS p75>0.1, AND speed_insights count > 50
• Billing dimension: speed-insights
• Scope: route
• Source citation: https://web.dev/articles/vitals

Routes where Core Web Vitals fall into Google's "Poor" band on real-user traffic. LCP > 2500ms, INP > 200ms, or CLS > 0.1 each hurt SEO and conversion. Surfaces one candidate per (route, metric) pair to keep recommendations focused.

---

external_api_slow

• Threshold: p75Ms > 2000 AND callCount >= 500
• Billing dimension: function-duration
• Scope: route
• Source citation: vercel-optimize gate threshold

External API hostnames with p75 latency above 2 seconds AND at least 500 calls in the window. External API latency is a primary driver of function duration cost when the upstream is on a hot path; a single slow stale call isn't worth recommending against.

---

isr_overrevalidation

• Threshold: writes/reads > 0.5 AND writes > 100
• Billing dimension: isr
• Scope: route
• Source citation: https://vercel.com/docs/incremental-static-regeneration

ISR routes with > 1 write per 2 reads. The revalidate interval is too aggressive relative to read traffic — many reads pay to regenerate. Investigate whether the page can tolerate a longer revalidate window or on-demand revalidation via revalidateTag.

---

middleware_heavy

• Threshold: middlewareInv/totalInv > 0.5 AND middlewareInv > 1000
• Billing dimension: edge-requests
• Scope: account
• Source citation: https://nextjs.org/docs/app/building-your-application/routing/middleware

Middleware invocations cover > 50% of total requests at non-trivial volume. The matcher is probably broader than necessary; narrow it to the paths that actually need auth/rewrites/headers.

---

observability_events_attribution

• Threshold: observabilityEventsShare > 0.20 (critical at > 0.30)
• Billing dimension: observability-events
• Scope: account
• Source citation: vercel-optimize gate threshold

Observability Events line item exceeds 20% of total billed cost. High share usually traces to low cache hit rate, middleware-heavy traffic, or unconstrained custom-span cardinality. No sampling lever exists for Observability Plus; reduce upstream invocations instead.

---

platform_bot_protection

• Threshold: botIdEnabled=false AND (botPct >= 0.05 OR edge_cost >= $25/window OR requests >= 14k/14d)
• Billing dimension: edge-requests
• Scope: account
• Source citation: vercel-optimize gate threshold

When BotID is disabled AND there is evidence (observed bot bandwidth share, edge cost, or substantial request volume) that bot traffic is non-trivial. Bot traffic inflates edge request counts without delivering user value; staged bot protection can reduce waste on bot-heavy projects. Skipped on quiet projects with no bot evidence — the recommendation would be noise.

---

platform_fluid_compute

• Threshold: fluid=false AND (any cold_start signal OR any route with p95>1000ms AND inv>1000)
• Billing dimension: function-duration
• Scope: account
• Source citation: vercel-optimize gate threshold

When Fluid Compute is disabled on a project that shows cold-start pressure (high cold-start rate) or sustained slow function p95 on hot routes. Fluid Compute reduces cold starts via instance reuse — recommend turning it on at the project level rather than per-route.

---

region_misconfig

• Threshold: single-region pin found AND routes.length > 20 (scanner-only branch)
• Billing dimension: function-duration
• Scope: account
• Source citation: vercel-optimize gate threshold

A single function region is pinned in vercel.json or per-route preferredRegion. Without per-region TTFB data (data gap), the gate can't quantify the geographic latency cost — but a single-region pin on a project with 20+ routes is worth auditing against Speed Insights traffic geo.

---

route_errors

• Threshold: count > 250 OR (totalRequests >= 1000 AND errorRate > 0.01)
• Billing dimension: function-duration
• Scope: route
• Source citation: vercel-optimize gate threshold

Routes producing > 250 5xx errors over the window, or with > 1% error rate on at least 1,000 total requests. Errored function invocations still bill at full duration; high error rates also poison user experience.

---

scanner-driven

• Threshold: per-kind: scanner matches.length >= threshold
• Billing dimension: mixed
• Scope: mixed
• Source citation: vercel-optimize gate threshold

Configured kinds emitted from scanner output. Each requires a minimum match count to avoid noise. Findings on cold-path or unmappable files are dropped unless the underlying scanner is trafficIndependent.

---

slow_route

• Threshold: (p95 > 500 AND inv >= 1400) OR (p95 > 1500 AND inv >= 250); disqualified when 5xx rate > 50%; Vercel Workflow runtime endpoints are hard-gated
• Billing dimension: function-duration
• Scope: route
• Source citation: vercel-optimize gate threshold

Routes with p95 function duration above 500ms at meaningful traffic (>=1,400 invocations in window), OR catastrophically slow routes (>1500ms p95 at any volume >=250). High duration drives both function-duration cost and user-perceived latency. Investigate sequential awaits, slow external APIs, missing caching, N+1 patterns. Routes with >50% 5xx rate are disqualified — those are reliability problems, not performance tuning targets, and surface via route_errors instead. Vercel Workflow runtime endpoints (/.well-known/workflow/v1/*) are hard-gated before launch because long-running step/flow requests are expected orchestration, not app-route bottlenecks.

---

uncached_route

• Threshold: requests > 500 AND hitRate < 0.5 AND getShare > 0.2 (missing getShare is gated)
• Billing dimension: edge-requests
• Scope: route
• Source citation: vercel-optimize gate threshold

Routes serving > 500 requests/period at < 50% cache hit AND at least 20% GET traffic. Each uncached GET request reaches the function, costing edge requests + function duration. Routes that are mostly POST/PUT/DELETE (Server Actions, mutations) are skipped — 0% cache is correct behavior there. Routes with missing method-share data are gated instead of launched. Auth-gated routes are disqualified separately.

---

usage_spike_triage

• Threshold: any-day total > 2x mean OR any-day SKU > 3x SKU mean
• Billing dimension: mixed
• Scope: account
• Source citation: vercel-optimize gate threshold

A single day in the billing window deviates sharply from the window baseline. Triage branches: bot or AI crawler spike, viral moment, pricing-model migration (legacy SKU → new), code regression. Without daily-granularity data, this gate stays dormant.

---

references/doctrine.md (verbatim)

Doctrine

The four non-negotiable rules that shape every action this skill takes. If a future change conflicts with one of these, the change is wrong.

Rule 1: Observability before investigation

The skill never reads a source file without an observability signal pointing at it. Step 1 (node scripts/collect-signals.mjs) is always first. Nothing reads source code until signals.json exists.

Why this fails when skipped: without metrics, the skill defaults to "grep the repo for known anti-patterns and complain." That produces noisy, low-impact recs that aren't tied to traffic, cost, or user pain. Metrics-first investigation keeps the skill focused on observed traffic, cost, and reliability signals.

Four-check first-pass (Enterprise)

When plan === 'enterprise', the gate run must surface these four checks before code-level recommendations. Field engineers confirm these are the highest-leverage account-level levers across every renewal audit:

1. Observability Plus enabled? From signals.observabilityPlus. If false, the whole audit degrades; surface as a top-of-report item.
2. Reverse proxy in front? Heuristic from response headers / CNAME chain (when collected). A non-Vercel CDN over Vercel ISR is usually a "dumb pipe" — wasted spend.
3. WAF rules enabled? From signals.project.security. BotID + managed rules absent on a project with bot evidence is the most common cost spike.
4. ISR read:write ratio. From metrics.isrReadsByRoute + metrics.isrWritesByRoute. Include CDN-tier reads (see data-collection.md) before flagging "writes > reads."

These checks anchor the Enterprise-tier report's opening narrative; code-level recs follow.

Rule 2: Deterministic gate before every sub-agent investigation

node scripts/gate-investigations.mjs is a pure-JS, LLM-free function. It reads signals.json and outputs {toLaunch, platform, gated}. Same input always produces byte-identical output (modulo appliedAt).

Every kind of candidate (uncached route, slow route, errors, cold starts, scanner findings, platform-level recs) has its threshold expression encoded as a gate(signals) → Candidate[] function in lib/gates/<kind>.mjs.

Failed gates surface in the final report, under "Not investigated in this run," with the exact reason they were held back. This is the user-facing trust mechanism: you see what we considered and chose to skip, and the reason.

Why this matters: the agent never decides "should I look at this route?" via LLM judgment. The threshold is mechanical. This eliminates the entire failure mode where the agent investigates routes it shouldn't (cold-path) and recommends fixes for routes that don't need them.

Rule 3: Candidate-bound investigation scope

When the gate emits a candidate with files: ['src/app/api/products/route.ts'], the agent reads ONLY that file (and its imports as the chain unfolds). It does NOT grep -r across the repo.

If you find yourself wanting to grep the whole codebase, stop and re-read the current candidate's question field. If the question doesn't constrain the search, the candidate is malformed — log it as gated and skip. Do NOT compensate with a wider search.

Why this matters: the agent's job is to verify and explain the metric anomaly the gate found, not to do a general code review. Wandering investigations produce drift, hallucination, and recommendations untied to the cost and performance data.

Scanner findings (the supplementary signal)

Static AST-grep scanners run in parallel with the metric-driven investigations. Their output is annotated with the per-file observability signal (function invocations: 1.2M; 95th percentile duration: 850ms; cache hit rate: 0% if the file maps to a hot route, COLD-PATH if it maps to a route with no traffic, NO-ROUTE-MAPPING if the file doesn't map to any route).

Default rule: scanner findings on COLD-PATH or NO-ROUTE-MAPPING files are dropped. They become recs only if the pattern is traffic-independent: build configuration, middleware matcher, source maps in production, raw script tags, React Compiler config. These don't care about traffic — they affect every request equally or affect the build itself.

The traffic-independent allow-list lives in each scanner's metadata.trafficIndependent: boolean field. Set it to true only when you can defend the claim.

Rule 4: Doc-grounded, version-aware recommendations — no hallucinations

Every recommendation must carry at least one citation from references/docs-library.json. Anything else is dropped at sanitizer time.

The library has two parts:

• URLs — Vercel docs, Next.js docs, SvelteKit docs, etc. Each declares applicableFrameworks (e.g., ["next@>=15.0.0"]).
• Cross-skill rule references — by name only (vercel-react-best-practices:async-parallel). The agent's host resolves these.

Three sanitizers enforce this:

• missing-citation — drops recs with empty citations[].
• unknown-citation — strips URLs not in the library, marks needsReview=true.
• version-mismatch — strips URLs whose applicableFrameworks doesn't match the project's framework@version (parsed from package.json).

Two verifier claim types check it: citation_in_library (URL ∈ allow-list) and citation_applies_to_version (semver match).

Why this matters: LLMs cite plausible-looking URLs that 404, or recommend Next 15 features to Next 13 users. Both are trust-killers. The allow-list closes the first failure mode; the applicableFrameworks field closes the second.

Performance citations cite observed data

Every performance claim cites the actual observability datum from signals.json — e.g., functionRoutes[/api/products].p95Ms=850. Estimated improvements are framed as ranges grounded in the observed baseline: "Reduce /api/products 95th percentile duration from 850ms toward ~250-400ms based on similar cached routes." Never an unanchored claim.

Cost framing is magnitude, never precise

Cost claims like $340/mo are forbidden. The dollar noise floor on projections is too high to justify precision. The impactMagnitude({currentCost, impactTier}) helper maps to phrases like "hundreds of dollars per month at current traffic" (computed against the user's actual vercel usage data).

The $-strip sanitizer enforces this at output time — any $N literal in customer-facing fields is stripped.

Performance numbers stay precise because they're observed, not extrapolated. We trust observed metrics; we don't trust dollar projections.

What good looks like

A good run produces:

• A small number (5-15) of recommendations.
• Every rec ties to a specific route or file plus a specific metric signal.
• Every rec carries before/after code and ≥1 citation matching the user's framework version.
• Cost framing uses magnitude phrases. Performance framing uses precise observed numbers.
• The "Not investigated in this run" section explains every other signal we saw and why we chose not to dig (cache hit rate was below threshold, 95th percentile duration was already healthy, etc.).
• No $N/mo strings, no fabricated URLs, no Next.js 15 features recommended to a Next.js 13 user.

What bad looks like (anti-patterns we will not ship)

• Recommendations from grepping the repo for known anti-patterns, without checking traffic.
• "Enable Fluid Compute" without a cold-start signal.
• "Add caching to /api/users" when the route has cookies() and is auth-gated.
• "Reduce the duration of /.well-known/workflow/v1/step" because a Workflow step is long-running. Workflow runtime endpoints are generated orchestration routes; high wall-clock duration there is expected unless a separate reliability/error signal points elsewhere.
• "Fix /api/chat/[id]/stream because it has high duration" without proving the stream does avoidable pre-first-byte work, high active CPU, duplicate invocations, or movable post-response work.
• "Save $340/mo by doing X" — invented precision.
• Citations to URLs that don't exist or that describe Next.js features the user's version doesn't have.
• Long lists of recs the user can't act on; every rec needs an evidence chain.

Out of scope

The skill is bounded to runtime cost and performance optimization on Vercel-hosted projects. The following are explicit non-goals; if signals or scanner findings surface in these areas, route them out:

• Deployment artifact size in isolation. Bundle size matters only when it shows up as runtime cost (cold start, FDT) or performance (LCP, INP). If the only effect is "the .next directory is large," it's not in scope.
• Build-time issues without runtime impact. Slow builds, build-cache misses, monorepo build fan-out — these only enter scope when they show up as Build Minutes billing pressure (then they go through the build-minutes-fanout gate). A 6-minute build that completes successfully and ships a small artifact is not a target.
• Security advisories and credential rotation. RCE in next-mdx-remote, leaked env vars, OIDC vs explicit-key auth hygiene — refer to a security skill, not this one. Exception: when a security setting is also a documented cost lever (BotID = bot traffic = edge cost), it enters via the platform_bot_protection gate.
• Commercial / billing-process trivia. Discount sliders, seat reconciliation, contract renewal mechanics. The skill can quantify which SKU is expensive; it does not negotiate.

references/observability-plus.md (verbatim)

Observability Plus Stop-And-Ask

Use this file only when signals.observabilityPlusBlocker is set. Do not silently continue into scanner-only mode unless the user chooses that path.

Why This Check Exists

This is a data dependency, not an upgrade pitch. The skill ranks work by observed route behavior so it can separate hot, expensive paths from code that only looks suspicious. These gates need per-route metrics:

Gate	Required signal
slow_route	Function duration and invocation count by route
uncached_route	Cache result and request count by route
cold_start	Function start type by route
route_errors	Function status by route
isr_overrevalidation	ISR reads and writes by route
middleware_heavy	Middleware invocations and duration
cwv_poor	Core Web Vitals by route
platform_bot_protection	Fast Data Transfer by bot category

Scanner-only mode can still catch traffic-independent code issues, but it cannot rank the hottest routes or prove cost impact. Make that tradeoff explicit before continuing.

User Template

Render this template first, then wait for the user's choice. Replace only <detail>. Do not add a preface; the heading is the opening line.

**Per-route metrics are unavailable.**

<detail>

This audit needs route-level metrics to rank fixes by observed latency, cache hit rate, error rate, cold-start rate, and Incremental Static Regeneration reads and writes. Without them, I can run a scanner-only audit for traffic-independent code issues, but I cannot tell which routes matter most or prove cost impact.

Docs: https://vercel.com/docs/observability/observability-plus

Choose one:
1. Enable Observability Plus, then re-run the metric-backed audit.
2. Continue in scanner-only mode for a limited audit.

If the host supports a structured question tool, use this exact customer-facing copy. Do not rewrite it.

{
  "question": "Enable Observability Plus and re-run, or continue with a limited scanner-only audit?",
  "header": "Observability Plus",
  "options": [
    {
      "label": "Enable and re-run",
      "description": "Use route-level metrics to rank the routes that matter most for cost and performance."
    },
    {
      "label": "Run scanner-only",
      "description": "Check traffic-independent code patterns without route ranking or proven cost impact."
    }
  ]
}

Use the full product name in this question. Do not abbreviate product names or metrics in customer-facing blocker copy.

After The User Chooses

If the user chooses Enable and re-run, stop after this short response:

Enable Observability Plus from the Vercel dashboard's Observability tab, then tell me to rerun. I'll restart the metric-backed audit once route-level metrics are available.

Do not include raw team IDs, org IDs, project IDs, pricing language, dashboard screenshots, or extra persuasion. The docs link in the blocker message already covers availability and billing details.

If the user chooses Run scanner-only, continue with the scanner-only steps below.

Blocker Copy

Blocker	Detail
payment_required	Detected: route-level metrics were recognized for this team, but these metric queries are not usable.
no_oplus_probe	Detected: this team does not expose the route-level metrics this audit needs.
not_linked	Detected: this app directory is not linked to a Vercel project.
forbidden	Detected: the Vercel CLI is authenticated to a team that cannot read this project.
project_not_found	Detected: the project ID is not visible to the authenticated team.
project_disabled	Detected: route-level metrics are enabled for the team but disabled for this project.
all_failed_other	Detected: every per-route metric query failed. Error code: <code>.

For not_linked, do not use the Observability Plus template. Link the app directory first:

vercel link --yes --project <project-name-or-id> --cwd <app-dir>

Add --team <team-id-or-slug> when the team is known. If the user supplied both app path and project name, run the link command instead of asking them what to do.

For forbidden and project_not_found, ask the user to confirm the exact Vercel project and team/personal scope before presenting the Observability Plus choice.

For project_disabled, do not present it as a team subscription problem. Ask the user to enable Observability Plus for this project, then re-run.

For no_traffic, do not use this template. Tell the user the project has no meaningful traffic in the 14-day window, then ask whether to run scanner-only mode now or come back after traffic accumulates.

Scanner-Only Mode

If the user picks scanner-only mode:

1. Re-run node scripts/collect-signals.mjs [projectId] --continue-without-observability > "$RUN_DIR/vercel-signals.json" 2> "$RUN_DIR/collect.stderr" if the current signals.json stopped at the fast blocker (usageError=NOT_COLLECTED_OBSERVABILITY_BLOCKED or project=null).
2. Run code scanners.
3. Launch only traffic-independent findings.
4. Render a clear data gap: per-route metric gates were skipped because Observability Plus data was unavailable.

Do not imply the scanner-only report is a complete optimization audit.

Back to vercel-labs/agent-skills (Vercel official skills) or Agent skills.