Dariusz Kowalski

@darco81.sdet.it

Context before LLM. Building AI tooling for test automation & developer workflow. → sdet.it 🇵🇱 PL/EN

Ran /doctor in Claude Code. 6,000 tokens gone before I typed anything. Just skill descriptions preloaded into the system prompt. I had ~80 skills. Did the math. Built a fix in a day. Thread on what I shipped 👇

Mac devs running local models - what's your HF cache size right now? du -sh ~/.cache/huggingface Mine: 92 GB. Most was Qwen3 variants I downloaded "to test", then forgot. Genuinely curious what others are sitting on.

Context before LLM is my brand line. Now there's a tool version: context before tokens, cleaners before swap. sdet-pulse. macOS + Claude Code observability and cleanup. - 22 optimization packs - HF/Docker/Xcode cleaners (mine was 92 GB) - -20% session cost on 65-session A/B pulse-test.sdet.it

Pulse landing page hero on dark CRT-style background. Cyan headline "Context before tokens. Cleaners before swap." Tagline: Active token optimization for Claude Code, system health for macOS. 22 packs that truncate, dedupe, route to Haiku, and nudge before tokens bleed. Plus monitoring, RAM relief, one-click cleanup on Mac. Status: Preview tier, open for testers. Counters: 22 packs, 21 actions, 14 tabs, 790 tests, 8 monitors, -20% session cost from author's own A/B on 65 sessions.

Hi 👋 New here. SDET pivoting into AI tooling builder. 🛠 What I'm shipping: - WCAG accessibility toolkit (5 AI specialists, OSS) - jarvis-brain (multi-tenant knowledge vault) - CDAT pattern (4-layer Playwright architecture) Context before LLM. Always. → sdet.it

sdet.it - AI engineer · architect · tooling builder

AI tooling, RAG deployments, QA training. AI agents for multi-tenant systems and test automation. 18 years shipping, AGPL distillates included.

sdet.it