Agent skill · software engineering · jeremylongshore
anth-load-scale
Implement load testing, auto-scaling, and capacity planning for Claude
Why this skill is useful
Includes executable scripts for load testing and capacity planning specific to the Claude API that extend the AI's capabilities.
What it needs
Requires anthropic, asyncio, python installed locally. Requires anthropic account access. About 3k tokens when loaded. Last updated 2026-08-07. 2,604 stars on the source repository.
What this skill does
Anthropic Load & Scale Overview Capacity planning and load testing for Claude API integrations. Key constraint: your rate limits (RPM/ITPM/OTPM) are the ceiling, not your infrastructure. Capacity Planning Load Testing Script Scaling Strategies Strategy When Implementation ---------- ------ --------------- Queue-based processing > 50 RPM sustained Redis/SQS queue + worker pool Model routing Mixed workloads Haiku for simple, Sonnet for complex Message Batches Offline processing 100K requests, 50% cheaper, no RPM impact Prompt caching Repeated system prompts 90% input token savings Request coalescing Duplicate prompts Cache identical request hashes Horizontal Scaling Pattern Error Handling Issue Cause Fix ------- ------- ----- 429 during load test Exceeded tier limits Reduce concurrency or upgrade tier Increasing latency under load Output queue saturation Reduce maxtokens Uneven request distribution No load balancing Use queue for fair distribution Prerequisites Confirm the organization/model rate limits, budget ceiling, test environment, concurrency cap, and success/latency/error thresholds before measuring capacity. Run only against an approved sandbox using synthetic prompts and a no-op result sink. Never stress production or use real customer content for load tests. Configure aggregate metrics and redaction: request counts, status classes, latency, token totals, queue depth, and 429 counts are sufficient; prompts, completions, keys, and tool arguments are not. Instructions 1. Calculate RPM, input tokens per minute, output tokens per minute, concurrency, and expected cost from the measured workload. Reserve headroom below provider and application limits. 2. Start with a small canary, then increase concurrency in bounded steps while a shared limiter coordinates all workers. Stop immediately at error, budget, data-scope, or latency thresholds. 3. Separate real-time traffic from batch work, and use queue backpressure rather than unbounded task creation. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills jeremylongshore/anth-load-scale