Percerno Skillsbeta
Measured, not claimed

Token benchmark: with and without Percerno Skills

Date: 2026-10-23. Model: the Claude model behind Claude Code on this machine (same for both arms). Each task ran in a fresh session, once per arm, hard-isolated:

Tokens in = input + cache writes + cache reads as reported by the CLI; cost as reported by the CLI. "ok" is a correctness check against known truth (the fixtures were generated, so the truth is exact).

Task Without Percerno With Percerno Verdict
Sum revenue by region in a 100,000-row CSV (2.3 MB) refused: "would very likely produce wrong numbers" · 9k tokens · $0.04 correct 4/4 · 36k tokens · $0.08 · 22 s impossible → done
Total of all amounts in a 3-page scanned invoice correct (read the page images) · 14k tokens · $0.08 · 17 s correct (OCR in sandbox) · 42k tokens · $0.08 · 65 s same cost, slower
Same for a 30-page scan batch correct · 120k tokens · $1.02 · 54 s correct · 166k tokens (mostly cheap cache reads) · $0.18 · 508 s 5.7× cheaper, slower
The 12 key figures in a 40-page Word report correct 12/12 · 61k tokens · $0.24 correct 12/12 · 44k tokens · $0.08 · 20 s 3× cheaper, faster
Build a 6-slide PowerPoint file no file; a Python script for the user to run · $0.07 file delivered · 405k tokens · $0.47 · 203 s · 16 tool steps impossible → done, but expensive
Write three LinkedIn posts delivered · 4k tokens · $0.03 delivered, run through the skill's post checker · 52k tokens · $0.17 5× more expensive

What this says

  1. The savings are real where the alternative is pushing file content through the model, and they grow with file size: 3× on a 40-page report, nearly 6× on a 30-page scan. Below a few pages the model reads the file as cheaply as we process it.
  2. Two of the six tasks are not "cheaper", they are "possible": the 100,000-row sum and the PowerPoint file. A chat assistant cannot do either. That is the product, not a token story.
  3. Text-only tasks cost more with the connector, by a factor of five here. The tool definitions add about a thousand tokens to every turn, a skill's instructions add 5–15k, and every extra tool step re-reads the whole context. For a LinkedIn post the skill buys method and a quality check, not savings. The server's instructions now tell the model to skip the lookup for such requests unless the user asks for a method.
  4. The deck is the harness problem in miniature. The pptx-generator skill prescribes exploration (five reference files, a QA pass), so the model took 16 steps and 400k tokens. With a skill that does the job in one command ("here is the outline as JSON, build it"), the same deck is one read_skill and one sandbox_run: roughly 40k tokens. That tenfold reduction is engineering we control, and it is the honest version of the "better harness, fewer tokens" claim.
  5. Wall-clock time is the box, not the design. Thirty pages of OCR took 8 minutes on the 2-vCPU instance; the model side was idle most of that time. The resize fixes it.

Reproduce

Runner, fixture generators and raw results are in docs/bench/ (run.py, make.py, make30.py, results.json); the fixtures are generated (make.py, make30.py) so the truth values are exact.

Try it on your own files. Free plan, no credit card.Get started free