Skill UI

This pattern addresses the fundamental limitation of context window size by treating the filesystem as a persistent, external memory. Agents can write large outputs, plans, or findings to files, offloading them from the active context. Instead of carrying all information, agents use targeted search and retrieval tools to load only the necessary context on demand, drastically reducing token bloat and improving performance over long, complex tasks.

A powerful tool leveraging OpenRouter and the Gemini 3 Pro Image model to create, modify, and composite high-quality images. Supports three primary use cases: prompt-only generation from text, detailed single-image edits (inpainting/outpainting), and complex multi-image composition. Users can specify resolutions up to 4K, making it suitable for professional creative and design workflows.

A continuous file-based planning system for multi-step AI agents. Maintains task_plan.md, findings.md, and progress.md files on disk, injecting project-specific planning context into lifecycle hooks. Auto-restoration reads only project planning files. Can inspect local agent session metadata with session-catchup.py --metadata, or replay limited excerpts with --replay. Optional controlled mode requests continuation only when supported by host. Never executes declared Markdown commands. No network data upload path. Used for research or work requiring 5+ tool calls.

Detects goal hijacking vulnerabilities where malicious instructions are hidden within external content sources processed by the agent, such as retrieved documents (RAG), uploaded files, or web page content. This skill simulates these attack vectors to ensure the agent only uses the provided context for the stated task and does not execute embedded hidden commands.

A four-tier memory store for AI assistants that promotes recurring facts from transcripts to project context and persona via deterministic recurrence gates, avoiding CLAUDE.md bloat and repeated corrections. Suitable for projects where agents need durable memory that is only committed after human adoption.

This skill enforces strict communication protocols for AI agents interacting in shared channels with external parties. It ensures agents speak only when explicitly addressed, prevents internal context leakage, and routes sensitive content to draft-only approval. Ideal for group chats, DMs, and business partnerships where privacy and security are critical.