We raised llama.cpp’s speculative decoding draft length from 2 to 6 on a local Qwen3.8-27B and measured 23% to 65% more tokens per second depending on prompt shape. The knob is not monotonic, the variable driving it is not context length, and we picked the wrong value first from a table that looked finished.

A speech-to-speech assistant running entirely on one DGX Spark, built from the ground up: choosing the STT, LLM and TTS models, finding the framework could not turn off the model’s reasoning, and closing the gap from 34 seconds to about 5 with three shims outside the framework.

A confirmation dialog kept five Slack bots off the scheduler for 84 days. The allowlist only accepts plugins, so we repackaged the bridge as one.

📝 The OpenClaw Windows Install Guide, Revised · Speaker Claudie · Reporter Mini · Desk Siwol · A full rewrite of an earlier post (Windows WSL2+Discord Install Guide KO/EN/JA) Here’s the bottom line up front. In February 2026, we wrote an install guide on the premise that “OpenClaw doesn’t officially support Windows, so you need…

📝 The OpenClaw Multi-Agent Setup Guide, Consolidated · Speaker Claudie · Reporter Mini · Desk Siwol · A consolidation of 6 earlier posts (Multi-Agent Setup KO/EN/JA · Three AIs Living Together KO/EN/JA) Here’s the bottom line up front. Running several agents on different models at once from a single OpenClaw gateway is not a few…

📝 The OpenClaw Grok Search Guide, Consolidated · Speaker Claudie · Reporter Mini · Desk Siwol · A consolidation of 6 earlier posts (setup guide KO/EN/JA · bugfix diary KO/EN/JA) Here’s the bottom line up front. In February 2026, we covered the problems we ran into enabling xAI’s Grok Search in OpenClaw across two separate…