ClawHub Skills That Actually Work: SFD Lab's Real Usage Report 2026

Who We Are and Why This Matters

SFD Lab currently runs three main projects: smallfiredragon.com, BuddyClaw, and BACAKU — with 14 AI agents handling content production, code development, security audits, and operations in parallel. We're heavy users, not casual experimenters. The skills we use, we use constantly.

The Skills Running Daily

self-improving-agent: The foundation of our error management system. When something goes wrong, the learning gets logged. When the same thing goes wrong twice, it becomes a rule. Our recurring error rate on well-defined tasks has dropped roughly 60% since deploying this skill.

edge-tts / voice-message: Morning briefings, alert escalation, async status updates — all audio. Once you've experienced the difference, text-only status updates feel like a downgrade.

humanizer: Last pass before anything goes to the CMS. The patterns it catches are subtle but real — the ones that make readers think "this was generated" even if they can't articulate why.

browser-automation: Weekly competitor monitoring, CMS verification after deployments, form submission for platforms without APIs. The natural language interface handles page variation that breaks traditional scrapers.

What We've Hit at Scale

API rate limits: the skills that make heavy API calls can hit rate limits when multiple agents use them simultaneously. We've had to implement request queuing for the high-frequency skills.

Context accumulation: skills with persistent memory can accumulate context that becomes stale. We run periodic context pruning to keep the working memory relevant.

The Bottom Line

The best skills are the ones that remove friction from things you do repeatedly. Install based on your actual workflows, not on what looks impressive in the description. And read the SKILL.md before installing — the quality of the documentation is a reliable proxy for the quality of the skill.