realistic-load-test

A Claude skill that finds out how many real users your website or API can handle.

Most load tests hammer one URL as fast as possible. That tells you a requests-per-second number, not how many people your app can serve. This skill makes every k6 virtual user behave like one of your actual users: the same pages, the same API calls, the same read/write mix, the same pauses to read and think. It then finds the highest number of those users your system serves while staying inside your latency and error criteria.

What it does

1. Scope. Confirms you own the target, picks staging vs production, sets criteria (default p95 < 500 ms, p99 < 1 s, errors < 1 %, or your SLOs), and estimates the cost.
2. Profile a real user. Reads your production logs (Nginx, Vercel, Cloudflare, ALB, JSON, CSV) or records a browser session, and works out requests per second per active user, endpoint mix, think time and journeys.
3. Write the virtual user. Generates a k6 journey.js with weighted journeys, real IDs, test accounts and safe handling of writes and third-party side effects.
4. Calibrate. A low-load run proves the virtual user matches the profile (rate within ±15 %, mix within a few points).
5. Load generators. Sizes and creates VMs (DigitalOcean, Hetzner, AWS, GCP, or Grafana Cloud k6), tunes them, and monitors them so the client is never the bottleneck.
6. Run. A fast staircase estimate, then 5-minute confirmation holds (median of 3) around it.
7. Report. The number, the per-endpoint latencies, what breaks first and the evidence for it, caveats, and teardown.

Install

Claude Code, as a plugin:

/plugin marketplace add arjaythedev/realistic-load-test
/plugin install realistic-load-test@realistic-load-test

Or clone the repo and copy skills/realistic-load-test/ into ~/.claude/skills/ (personal) or .claude/skills/ (one project).

Then ask Claude something like "how many users can my app at staging.example.com handle?".

Requirements

• k6 locally (for smoke tests and calibration)
• Python 3.9+ (profile scripts use only the standard library)
• Optional: Node (for npx playwright browser recording), a cloud CLI (doctl, hcloud, aws, gcloud)

Only test what you own

Load testing a system you don't control is indistinguishable from a denial-of-service attack. The skill asks you to confirm ownership and refuses obvious third-party targets. It also keeps load away from third-party APIs your app calls, and warns about firewall, rate-limit and billing effects on your hosting platform.

Layout

skills/realistic-load-test/
  SKILL.md                 the workflow
  references/              profiling, journey writing, criteria, load generators, platform notes, report template
  scripts/                 analyze_logs.py, har_to_profile.py, compare_profile.py, setup-loadgen.sh,
                           loadgen-monitor.sh, run-holds.sh
  templates/               lib.js, hold.js, staircase.js, journey.example.js (k6)

evals/ holds the test cases used to develop the skill (runnable with claude plugin eval); they are not part of the installed skill.

The method (closed-model users with think time, staircase with per-step criteria, median-of-3 holds, generator monitoring) comes from the $12 server benchmark series on benchmarks.arjaythedev.com.