Benchmark results

Compress performance findings without detaching numbers from conditions.

Condense supplied benchmark results for a reader comparing measured behavior. Preserve the workload and measurement limits that determine whether the comparison is useful.

State what was compared, under which conditions, and which metric matters. Keep units, sample counts, dataset size, hardware, concurrency, warm-up treatment, and uncertainty when supplied and relevant. Pair each headline result with its baseline. Distinguish latency percentiles from averages, throughput from latency, and measured results from estimates.

Remove repeated descriptions of the harness and commentary that adds no evidence. Use a small table when it makes like-for-like measurements easier to compare. Do not combine incompatible runs, derive an unsupported aggregate, or label a result significant without a supplied analysis.

End with the practical conclusion supported by the data and its strongest limitation. A faster result on one workload is not a general speedup. Do not run benchmarks, invent missing measurements, or quietly omit a regression.

Example

Parser benchmark

Before 105 words

We compared parser A with parser B on the same 8-core machine using a fixed set of 10,000 JSON documents. Each document was 2 KB, and concurrency was 1. We made five measured runs per parser after one warm-up run, which we excluded. Parser A had a median run time of 420 ms and a peak memory use of 82 MB. Parser B had a median run time of 310 ms and a peak memory use of 109 MB. This workload favors B on time but A on memory. We did not test larger documents or concurrent requests, and we did not calculate confidence intervals.

After 62 words

Fixed workload: 10,000 × 2 KB JSON documents; same 8-core machine; concurrency 1. Five measured runs per parser after one excluded warm-up. Parser | Median run time | Peak memory A | 420 ms | 82 MB B | 310 ms | 109 MB B was faster; A used less memory. Larger documents and concurrent requests were untested. No confidence intervals calculated.

Examples illustrate the method. They do not measure model output.

Files and sources

prompt.txt Complete prompt as plain text.
SKILL.md The same prompt with a skill header.
llms.txt (JSON) Prompt, metadata, example, and sources.
LICENSE MIT license text.

An original sho.rten.it skill. Source notes.