Dotclaudefiles

Search, evaluate and select libraries when a problem is "standard" — meaning someone has almost certainly solved it already. The strongest trigger is mid-planning of a new feature — while drafting or reviewing an implementation plan (plan mode, task-planning), run the standard-problem test on each planned component before the plan commits to a manual implementation. Also use this skill whenever the user is about to implement non-trivial functionality that sounds like a solved problem (programmatic SSH, retry/backoff, argument parsing, schema validation, date handling, HTTP routing, queues, templating), asks "should I implement X myself or use a lib?", "is there a library for X?", asks to compare libraries, or starts writing infrastructure-flavored code from scratch. Also use it when evaluating whether an existing dependency should be replaced, or when designing a public API and the prose-quality taxonomy (Facade / Fluent Interface / Internal DSL) applies.

Constraints, structure and template for implementation plans. Invoke immediately upon entering plan mode, before drafting any plan — and also when reviewing an existing plan, re-planning mid-task after a scope change, or when the user asks for an implementation plan, spec, or breakdown in any form. The plan is not ready to present until it satisfies this skill's checklist.

Designs, writes, or audits dynamic workflows for Claude Code — the JavaScript orchestration scripts the Workflow tool runs to fan many subagents across one large job (triggered by 'ultracode' or by asking to "use a workflow"). Use this skill proactively whenever the user says "create a workflow", "write a dynamic workflow", "orchestrate subagents", "fan out agents", "parallelize this with agents", "codebase-wide audit", "massive bug hunt", "ultracode workflow", "port this from X to Y", "migrate the whole repo", or describes work a single pass can't hold (repo-wide audits, large migrations, language ports, verified multi-source research). Also use it to audit or review an existing workflow script — "audit this workflow", "review my orchestration script", "is this workflow correct", "improve this workflow" — checking data-flow shape, per-role thinking load, and runtime safety. For authoring a single isolated sub-agent that reports back once, this is not the skill. Reach for this even when the user only hints at

Crea archivos markdown de sub-agentes de Claude Code (con YAML frontmatter y system prompt) listos para colocar en ~/.claude/agents/, .claude/agents/ o agents/ de un plugin. Usa esta skill proactivamente cuando el usuario diga "crea un agente", "create an agent", "necesito un sub-agente que...", "scaffold an agent", "haz un auditor", "make a code reviewer agent", "agente que revise", "agente que audite", "agente que investigue", o cuando describa funcionalidad especializada que conviene aislar en un sub-agente (ej. "necesito algo que revise PRs", "quiero un agente que analice seguridad"). Cubre seleccion de tools con principio de menor privilegio, tuning de description para auto-invocacion, eleccion de arquetipo (auditor / researcher / implementer / orchestrator), calibracion de modelo + effort segun la carga de razonamiento de la tarea (mas mecanico = modelo mas ligero y menos effort) e incluye un sistema de scoring de confianza 0-100 con umbral >=80 para agentes tipo auditor.