inference-optimisation (Sybil Solutions)
A skill that guides an agent through speeding up a local LLM server step by step: work out the hardware ceiling, change one setting at a time, measure against controls, log every result, and only keep changes that pass a long-context stability check. Covers decode, speculative decoding, prefill, KV cache and startup time. MIT.
Test results
Run with our skill harness: the skill alone in a throwaway project, invoked by name, and the real output file checked. Skills are tested in Claude Code and Cursor only so far.
| Client | Result | Last passed | What we checked |
|---|---|---|---|
| Claude Code | Not tested yet | never | |
| Cursor 2026.11.17-f532c45 | Passed Tested 2026-11-17 | 2026-11-17 | inference-optimisation: {'THEORY.md with token': True, 'decode ceiling at 8K ≈ 38 tok/s (220 GB/s ÷ ~5.82 GB)': True, 'prefill ceiling at 8K ≈ 1,650 tok/s (50% of 60 TFLOPS)': True, 'says 80 tok/s target is above the cei |
Skill version: github.com/sybil-solutions/inference-optimisation-skill @ fda968e7 (2026-11-16).
Tested on ToolPassed badge
Maker of inference-optimisation (Sybil Solutions)? It passed in Cursor on 2026-11-17. Add this free badge to your README; it links back to this listing. Free, no strings: the badge never changes a result.
[](https://toolpassed.com/skills/inference-optimisation-skill/)Get it
The skill’s files and install instructions are in its repository: github.com/sybil-solutions/inference-optimisation-skill.
Not working?
Fixes from our own test runs: Skill not working in Cursor?