Loading...
Loading...
Found 1,211 Skills
[WHAT] Universal content intake system for URLs (GitHub repos, YouTube videos, articles, PDFs) and skill packages (skills.sh, skill:// protocol) [HOW] Phase 1: Clone repos/fetch transcripts/scrape content/resolve skills to ~/lev/workshop/intake/. Phase 2-3: Load workshop/intake.md for full analysis [WHEN] Use when user provides a URL to analyze, says "intake/download", wants to evaluate external content, or references a skill package [WHY] Systematically evaluates external content and skill packages for adoption/adaptation with tier classification and ADR creation Triggers: "intake", "download", "analyze this url", "check out this repo", "review this video", "evaluate content", "install skill", "skill://"
Use when users provide vague, underspecified, or unclear requests where they need help defining WHAT they actually want - across ANY domain (writing, analysis, code, documentation, proposals, reports, presentations, creative work). Trigger aggressively when users express VAGUE GOALS ("make this better", "improve our X", "figure out what to include", "I don't know where to start", "kinda lost on what to do", "not sure what this means"), UNDEFINED SUCCESS ("should look professional", "explain this clearly", "make it convincing", "whatever works best", missing constraints/audience/format), COMMUNICATION UNCLEAR ("how do I explain/communicate this", "my team gets confused when I describe it", "help me figure out what to ask about X"), AMBIGUOUS REQUIREMENTS ("analyze the data" without saying what to look for, "improve documentation" without saying how, "make it more robust" without defining robustness, any request with multiple valid interpretations), or META-PROMPTING ("optimize this prompt", "improve my prompt", "make this clearer", "review my instructions", learning about prompt frameworks like CO-STAR/RISEN/RODES, understanding what makes prompts effective). Trigger for non-technical users and ANY situation where the request needs refinement, structure, or clarification before execution can begin. When in doubt about whether a request is clear enough - trigger.
Analyse agent execution to find wasted tool calls, wrong turns, and blind alleys. Optimise agents to reach their goal in the fewest turns, tokens, and least time. Recommend harness/model changes — never apply without user approval.
Use when managing AI Hub account, API keys, balance, usage, or API endpoints. Use when user says "AI Hub", "add AI credits", "create API key", "check AI usage", "auto-recharge", "AI Hub endpoint", "AI Hub base URL", "how to use AI Hub API", "LLM API", "AI API", "OpenAI compatible", "Anthropic API", "GPT", "Claude", "Gemini", "DeepSeek", or "Grok" in the context of Zeabur.
Use when starting a new project with Maestro or when no .maestro.md context file exists yet. Run once per project.
Fact-forcing gate that blocks Edit/Write/Bash (including MultiEdit) and demands concrete investigation (importers, data schemas, user instruction) before allowing the action. Measurably improves output quality by +2.25 points vs ungated agents.
Create ShinkaEvolve task scaffolds from a target directory and task description, producing `evaluate.py` and `initial.<ext>` (multi-language). Use when asked to set up new ShinkaEvolve tasks, evaluation harnesses, or baseline programs for ShinkaEvolve.
Use when assessing AI/ML systems for prompt injection, jailbreak vulnerabilities, model inversion risk, data poisoning exposure, or agent tool abuse. Covers MITRE ATLAS technique mapping, injection signature detection, and adversarial robustness scoring.
Expert guidance on AI Agent Harness architecture based on the comprehensive Claude Code analysis book
Expert knowledge of agentic AI design patterns for autonomous agent development
A curated collection of research papers and resources on agentic reasoning for Large Language Models, organized by planning, tool use, search, self-evolution, and multi-agent systems.
Integrate TileGym kernels into Hugging Face `transformers` models by replacing the library's submodule(s) and certain class(es)' implementations, and patching certain class(es)' init/forward/load weight methods prior to instantiating models. Used when the user requires integrating TileGym kernels into `transformers` models.