skillhub — The Skill Composer for AI Agents
There are 50+ tools to install AI agent skills. There is exactly one that merges them.
Compose skills from 13 ecosystems — Anthropic, OpenAI, GitHub Copilot, Microsoft, Google, Vercel, and 8 more — into one unified expert skill with AI-powered conflict resolution.
pip install skillhub-ai
# Merge skills from 3 different companies into one expert
skillhub compose anthropic:claude-api openai:aspnet-core google:agent-platform-deploy -o cloud-expert
# AI-powered merge — Claude resolves section conflicts intelligently
skillhub compose python-patterns security-review -o secure-python --strategy aiSupported ecosystems: anthropic: · openai: · copilot: · microsoft: · google: · skills.sh: · agency-agents: · addyosmani: · scientific: · antigravity: · gamedev: · tech-leads: · github:
Fragile Safety: Automated Circuit Discovery is Vulnerable to Dormant Feature Bundling Chandrashekar DP — Zenodo Preprint, June 2026 · arXiv submission pending
Automated circuit discovery tools used for AI safety verification are structurally vulnerable to adversarial inputs. A linear probe achieving 97.5% accuracy on clean data fails completely on adversarial inputs (0% detection rate). The adversarial distribution is geometrically inseparable from clean positives at the anchor token (cosine similarity 0.989). Reproduced on Pythia-410m using TransformerLens. Practical mitigation: context-aware probing at the last sequence token achieves 100% adversarial detection with clean accuracy preserved.
📄 Preprint — Zenodo · 💻 Code
| Project | What it is | Status |
|---|---|---|
| skillhub | Skill composer for AI agents — merge skills from 13 ecosystems | ✅ Live on PyPI |
| TransformerLens | Contributing test coverage for Google DeepMind's mech-interp library (3,500+ ⭐) — 4 PRs merged | ✅ Ongoing |
| enterprise-rag-patterns | Production RAG architectures from the GenAI in Production newsletter — RCA agent (80% triage reduction), RAG eval at scale | ✅ Active |
| ComplianceShield | AI compliance assistant — PII detection, session audit logging, HITL review, multi-LLM | 🔨 Building |
| nanoGPT → micrograd → minbpe | Working through Karpathy's series — transformers from raw math, not API calls | 📖 Ongoing |
The LLM observability gap — LLMs are in production everywhere. Engineers have nothing equivalent to what researchers have in TransformerLens. No production-grade "Sentry for model reasoning." This is the problem I keep returning to.
Evaluation is harder than training — Writing the RAG evaluation framework made this concrete. Measuring whether an LLM is correct is a deeper problem than making it correct. Most teams skip it. The ones who don't, ship reliable AI.
Interpretability as infrastructure — The people who build the measurement tools shape what the whole field builds next. Open-source research infrastructure is underrated leverage.
| Book | Why | |
|---|---|---|
| 🧠 | Concrete Problems in AI Safety — Amodei et al. | Foundation for safety research |
| 📐 | Attention Is All You Need — Vaswani et al. | Primary source, always worth re-reading |
| 🏗️ | Designing Machine Learning Systems — Chip Huyen | Production ML thinking |
| 🧩 | Thinking, Fast and Slow — Kahneman | Applies directly to how humans evaluate AI outputs |
"Build it end-to-end. Ship it. Then explain why it works."
