Clean in code review.
Slow on a cheap Android at 4G.

One install gives your agent a measured pre-launch performance fix: a production-mode baseline, four read-only subagents auditing twenty known slow-downs across images, bundles, database queries and servers, then one fix pass — proved with the same mobile Lighthouse run, three times, before and after.

4 read-only subagents 20 audited slow-downs 3 runs per page, median 0 features removed

for coding agents · free, open, no auth

install

Copy one line. Ship with numbers.

Detects Claude Code, Cursor, Codex, Qoder and more, then writes the skill into the right folder. Add -l to see what the repo ships before anything is installed.

npx skills add arm-pz/pre-ship-performance-audit

Quality record

Know what you are installing.

Version, date and limits are read from the repo on every sync, and the check runs every night. Real-world runs are listed only with their evidence.

Version
Unversionedcommit 748a3b5
Updated
24 Oct 2026last commit to the repo
Works in
Claude Code, Codex, Cursor, Qoder and any agent that loads a SKILL.mdplain text; no API keys
Nightly check
PassingNo failures or warnings. How it is checked
Last real-world run
No public run yetRuns are published only with their evidence.
Known limits
Not built for:
  • Authorization and abuse testing (use pre-ship-security-red-team) or exposed secrets and source files (use web-preflight)

the four groups

One subagent per lane. None of them edits.

The subagents inspect and report — file, route, evidence, proposed fix. The main agent implements centrally, after the findings are prioritized into a P0–P3 backlog.

Images, fonts, third-party 1

The bytes most pages actually drown in.

  • WebP/AVIF at rendered size
  • framework image component
  • hero prioritized, rest lazy
  • self-hosted fonts, one or two
  • unused scripts cut, widgets deferred

Frontend & rendering 2

Ship less code, render it earlier.

  • bundle analyzer on big deps
  • whole-library imports trimmed
  • charts/editors/maps deferred
  • homepage data fetched server-side
  • AI answers streamed
  • skeletons without layout shift

Database & APIs 3 · high effort

The one group that must read query plans, not folklore.

  • indexes for filters, sorts, RLS
  • (select auth.uid()) per policy
  • N+1 collapsed to one query
  • keyset pagination
  • public-only shared caches
  • heavy work queued

Server & network 4

Where the request actually travels.

  • functions near the database
  • static / ISR for calm pages
  • sequential calls parallelized
  • low-end Android, slow 4G walk
  • compression, headers, redirects

The measurement contract all

Before any fix: production build or approved preview, mobile Lighthouse ×3 on the homepage and the slowest pages, medians recorded with exact commands. After: the same URLs, the same configuration, the same device. No comparable conditions, no percentage claim.

  • median of 3, never best-of
  • before/after table delivered
  • per-fix table with measured impact
  • unrun substitutes labeled

house rules

What it refuses to do.

Features are not fuel for a score

The audit fixes slowness by removing waste, never functionality. If a speedup could only come from dropping a feature or a dependency, that is a question for you, with expected benefit and trade-offs written out first.

No baseline, no claim

If a production build or preview cannot run, the skill says so instead of inventing numbers. Substitutes are labeled. And "this is a best practice" never counts as evidence of your bottleneck — the file, the route and the measurement do.

Approvals at the blast radius

Removing dependencies, changing deployment regions, or touching database schema, indexes, RLS or settings stops for your go-ahead. And a user-specific response is never pushed into a shared cache — speed without data leaks.

Ship the fast version, with the numbers to prove it.

Free, open, and it runs entirely inside the agent you already use.

Install pre-ship-performance-audit Read the source