A living, self-improving cognitive architecture that thinks, learns, remembers, adapts, and evolves — permanently. Use for ANY task where quality must compound over time: code, analysis, writing, debugging, research, planning, design, or multi-step workflows. ALWAYS trigger when: the task feels familiar, mistakes must not repeat, quality must exceed previous work, the user signals frustration ("you always get this wrong", "remember last time", "learn from this"), or when the task is complex enough that experience should shape the approach. This is not a skill. It is a mind that grows with every task it touches. The longer it runs, the smarter it becomes.
A self-review skill that attacks Mirage's own output from four adversarial perspectives — security researcher, failure analyst, scale engineer, and future maintainer — before presenting it to the developer. Also enforces explicit uncertainty quantification on every claim: distinguishing what is verified, inferred, or hypothetical. ALWAYS activate after any implementation of: authentication, payment logic, data mutations, file handling, external API calls, or any feature the developer has flagged as high-stakes. Also activate when the developer asks "is this safe?", "will this scale?", "what could go wrong?", or "are you sure about this?" This skill is the difference between Mirage presenting its first draft and presenting a production-ready implementation that has already been internally stress-tested.
A deep structural intelligence layer that builds and queries a living model of any codebase — its dependency graph, semantic index, git history, defect density, coupling patterns, and hidden invariants. ALWAYS activate when: starting work on an unfamiliar codebase, before any refactoring or architectural change, before touching files that have a history of instability, when asked "what will break if I change X", or when the task requires understanding how parts of the system relate to each other. This skill transforms Mirage from an agent that reads files to one that understands systems. It is the structural memory layer — where adaptive-mind is episodic memory of what happened, codebase-archaeologist is semantic memory of what exists and how it connects.
Use when an agent needs to search the web, find images/news/videos/books, or extract content from a URL. Covers the ddgs Python library, CLI, and MCP server integration.
A structured reasoning skill that generates competing hypotheses, designs falsification tests, eliminates candidates with evidence, and — for high-stakes architectural decisions — runs a multi-perspective debate where each perspective actively tries to break the proposal. ALWAYS activate for: any bug where the root cause is not immediately obvious, any architectural decision that is difficult to reverse, any situation where the first explanation feels too convenient, or when asked "why is this happening" or "which approach should we take". This skill is the antidote to the Einstellung Effect — it prevents the most familiar solution from blocking the best one. It makes Mirage reason like a scientist and decide like a senior engineering committee.
The teaching and knowledge-deepening layer of the adaptive-mind system. Activates when Claude needs to explain a concept just used, teach another instance, diagnose why a mistake keeps recurring, articulate a pattern so it becomes permanently understood, generate reusable teaching material from experience, or detect where knowledge is held but not understood. ALWAYS activate in combination with adaptive-mind. ALWAYS activate when the user says "explain why", "teach me", "why do I keep doing this", "what's the right way", "help me understand", "what would you tell a junior developer", or "how would you explain this to someone else". The Feynman principle governs: teaching is the fastest path from held knowledge to genuine understanding. The more Claude teaches, the deeper it understands. This skill is not supplementary — it is the amplification layer that makes adaptive-mind compound faster.