Turns recurring work into reusable skills. Use when a task being done now has been done before in this project, when the user says "make this repeatable", "turn this into a skill", "we'll be doing this again", or at the natural end of a multi-step task that will clearly recur. The rule of thumb - do it once by hand, codify it the second time.
Self-corrects skill and instruction files after a genuine instruction defect is confirmed. Use when Claude errors during skill execution and finds the instruction was at fault, or when the user reports a mistake AND verification confirms the instruction (not the execution) caused it. Applies succinct fixes so a real defect never repeats — without mangling correct instructions on a false alarm.
Creates high-quality Claude skills following official best practices. Use when the user asks to create a new skill, improve an existing skill, or needs guidance on skill authoring. Includes guided scoping, proper structure, naming conventions, progressive disclosure patterns, current frontmatter options, and quality validation.
Conducts in-depth analysis of a specific source or topic, producing a comprehensive, information-dense summary for research synthesis. Use when detailed analysis and documentation of an individual source is needed as part of a larger research effort. Primarily used by the research-analyst agent inside the deep-research workflow.
Conducts comprehensive multi-source research on a topic - clarifies scope, plans research angles, spawns parallel analyst subagents to study each source in depth, and synthesizes findings into cited documentation. Use when the user wants thorough research, a literature review, a competitive scan, or says "deep research", "research this properly", "build me a research doc".
Read-only audit of what Claude Code can currently see and do - settings layers, permission rules and modes, connected MCP servers, hooks, additionalDirectories, installed plugins, and sensitive files not covered by deny rules. Use when the user asks "what can you access", "audit my setup", "is anything exposed", or before starting work with confidential material. Makes no changes; never echoes secret or credential values.
Hardens Claude Code permissions with safe defaults for knowledge work. Use when the user wants to protect secrets or sensitive files, control what runs without asking, set up guardrails before working with confidential material, or says "secure my setup", "lock this down", "set up permissions". Writes settings.json rules the harness enforces - stronger than any prompt instruction.
Captures durable learnings into a second-brain knowledge base (a folder with a conventions.md or KB-declaring CLAUDE.md) when conditions warrant - a reusable insight, a framework crystallizing, a constraint discovered, a decision trail completed. Triggers on "save this to my notes/knowledge base", "file that under", "worth keeping", or the natural pause at the end of substantive work where a real learning emerged. Only operates when a knowledge base exists; not for process chatter, debugging logs, or tactical one-offs.