Alfonso-Castano

Capture implementation decisions for a feature before planning begins — libraries, error handling, edge cases, scope boundary. First step of the full feature loop.

Run the task files produced by /feature-plan, one fresh-context executor subagent per task (or wave), one commit per completed task. Third step of the full feature loop.

Research (conditionally) and decompose an approved feature into per-task files. Second step of the full feature loop. Supports --thorough for a fuller research pass on unfamiliar or high-risk work.

Fast path for a small, well-understood feature or fix. Skips discuss/plan/verify ceremony but still writes one task file and still requires evidence before claiming completion.

Entry point for building a feature. Classifies scope and routes to the fast path or the full discuss/plan/execute/verify loop. Also reports status on an in-flight feature.

Single merged review pass (spec compliance + code quality + goal alignment) plus a hard evidence gate before the feature can be marked done. Fourth step of the full feature loop.

Initialize a new project from a rough idea. Deep interrogation, optional domain research, and a roadmap if needed — writes durable context to .context/ that every future session and skill reads from. One-time, whole-project setup — not for individual features within an already-initialized project.

Generate self-contained, agent-ready prompts for delegating work to fresh Claude sessions or sub-agents. Use whenever you need to hand off a task (fix, feature, debug, refactor) to another agent that has no prior context. Also use when crafting sub-agent prompts within workflows, or when any skill/workflow needs a high-quality delegation prompt. Trigger on: "generate a prompt for", "write a prompt to", "hand this off to another agent", "create a fix prompt", or any context where work must be delegated to a fresh agent with zero shared context.

Multi-agent strategy skill for deciding how to parallelize work and generating ready-to-use spawn prompts. Invoke this skill — automatically or via /orchestrate — whenever a task scope feels too large for a single context window, when specialization would clearly benefit different parts of the task, when independent parallel tracks exist, or when the complexity suggests the main context would be polluted by keeping everything inline. This skill diagnoses whether to use a single session, subagents, or an agent team; decomposes work into tracks with explicit file ownership; seeds each agent's context; and produces ready-to-use spawn prompts. Subagents are the default parallelization tool — use them liberally. Agent teams have a high bar. When in doubt about whether to parallelize, invoke this skill to decide.

Generate self-contained, agent-ready prompts for delegating work to fresh Claude sessions or sub-agents. Use whenever you need to hand off a task (fix, feature, debug, refactor) to another agent that has no prior context. Also use when crafting sub-agent prompts within workflows, or when any skill/workflow needs a high-quality delegation prompt. Trigger on: "generate a prompt for", "write a prompt to", "hand this off to another agent", "create a fix prompt", or any context where work must be delegated to a fresh agent with zero shared context.

Educational assistant for ML Financial Risk System. Teaches ML theory, PyTorch, Docker, architecture decisions, and AI engineering principles with adaptive depth control. Use when user asks to learn about or understand project concepts. Supports --level flag (beginner/intermediate/advanced).