21 October 2026

Multi-phase security audit skill for coding agents

A skill for coding agents that runs a security audit in six phases using isolated agents. Findings are written to findings.json and re-verified by a fresh agent.

• Phases: reconnaissance, hunting, candidate validation, structured output, independent verification, report.
• Each candidate finding is checked by a fresh verifier that tries to refute it.
• It is the starting point of Cloudflare's vulnerability discovery harness.

An agent skill that writes as little code as possible

A skill that makes AI agents reuse existing code before writing new code. A benchmark on Claude Code records 53% less code.

• Date picker example: 335 lines without Ponytail, 10 lines with Ponytail.
• /ponytail-review and /ponytail-audit check for bugs, security, load, and missing tests.
• Validation, security, and accessibility are not cut.

Turn a codebase into a queryable knowledge graph

Type /graphify . in an AI assistant and the project is mapped into a graph. Code is parsed locally with the tree-sitter AST, with no LLM and no vector store.

• Each edge is tagged EXTRACTED (present in the source) or INFERRED (deduced by graphify).
• Output: graph.html, GRAPH_REPORT.md, and graph.json.
• graphify query and graphify path A B traverse the graph.

A collection of ready-to-install specialist AI agents

A collection of AI agent personas. Each agent has its own expertise, workflow, and deliverables. They can be installed into Claude Code, Cursor, Codex, Gemini CLI, and more.

• Install by division, for example --division engineering,security.
• There is a desktop app for macOS, Linux, and Windows that installs agents with a click.
• Each agent file contains an identity, a mission, code examples, and success metrics.

Role-based plugins for Claude Cowork

An open source repo with 11 plugins for Claude Cowork, also compatible with Claude Code. Each plugin contains skills, connectors, slash commands, and sub-agents for one job function.

• Example plugins: productivity, sales, customer-support, product-management.
• Can be customized to a company's tools, terminology, and processes.

One plugin marketplace for many coding tools

A catalog of 94 plugins for Claude Code, Codex, Cursor, OpenCode, GitHub Copilot, Antigravity, and Pi. The components share the same Markdown source.

• The local source contains 202 agents, 185 skills, and 105 commands.
• A single skill can be installed on its own with npx skills add wshobson/agents --skill python-testing-patterns.

A local skill catalog for AI agents

A library of 2,724+ SKILL.md playbooks. AAS Core helps Codex or Claude search the catalog, record the skills the agent chose, and show a plan before any target is changed.

• Core does not rank or recommend skills. The agent chooses.
• Apply and recovery are still experimental.
• A CLI, a local MCP server, a plugin, and a Workbench are available.

A foundation model for financial market candlesticks

A family of decoder-only Transformer models trained on K-line data from more than 45 exchanges. A tokenizer converts OHLCV data into hierarchical discrete tokens.

• Four sizes: mini 4.1M, small 24.7M, base 102.3M, large 499.2M parameters.
• The mini, small, and base weights are on Hugging Face. Large is not open source.
• The README notes acceptance at AAAI 2026.

A fast LLM inference and serving engine

An LLM inference and serving library that originated at the Sky Computing Lab, UC Berkeley. It uses PagedAttention to manage attention key and value memory.

• Continuous batching, chunked prefill, and prefix caching.
• FP8, INT8, INT4, GPTQ/AWQ, and GGUF quantization.
• Speculative decoding and CUDA/HIP graphs.