pricing-prediction

Ejecutar experimentacion notebook-first para modelos tabulares de machine learning y deep learning. Usar cuando Codex deba crear un notebook en `notebook/`, cargar un dataset, hacer feature engineering extenso, comparar boosters con predominancia en CatBoost, probar redes neuronales en PyTorch con `mps`, validar con 5 folds, reportar siempre las metricas elegidas y declarar el modelo ganador de cada corrida o nueva iteracion.

Build and debug browser-based web scrapers with a Selenium-first workflow and agent-browser-assisted site discovery. Use when Codex needs to inspect dynamic websites, dismiss cookie banners, identify stable selectors, paginate listing pages, extract product cards or PDP fields, save JSON or CSV outputs, or prototype repeatable scrapers for marketplaces such as falabella.com.pe.

Execute implementation work through a branch-first GitHub pull request workflow with a mandatory self-review gate. Use when Codex is implementing a feature, bug fix, refactor, or non-trivial change and must create a fresh branch, open a detailed pull request, publish an exhaustive severity-ranked code review with remediation options, and post a final GO or NO-GO verdict comment on the PR.

Implement and modify backend systems with Flask. Use when Codex is asked to build or change APIs, backend endpoints, request handlers, business logic, server-side integrations, authentication, persistence, jobs, admin services, or any backend application behavior. Default new backend work to Flask and Flask blueprints. Also applies to Spanish prompts like "backend", "api", "endpoint", "ruta", "controlador", "servicio", "modelo", "cambio en backend", or "implementacion backend". If the repository already has a different established backend framework, inspect it first and avoid mixing frameworks without calling out the conflict.

Leer comentarios y code reviews de un pull request, analizarlos de punta a punta, responder cada observacion y levantarla con cambios minimos y poco invasivos. Usar cuando el usuario pida revisar feedback de un PR, responder comentarios de GitHub, implementar observaciones sin cambiar la logica existente, o cerrar hilos de review de forma segura.

Analyze a git worktree, group changed files by cohesive change intent, and propose an approval-ready Conventional Commits plan before any staging or commit happens. Use when Codex needs to split a dirty worktree into logical commits, infer commit type and scope, review mixed changes, or prepare conventional commit messages for explicit user approval.

Crear planes de implementacion detallados para features y guardarlos como archivos Markdown en /plans. Usar cuando Codex deba implementar una funcionalidad nueva, definir alcance tecnico, mapear archivos a crear o modificar, listar pruebas, o preparar validaciones con Ruff y mypy antes de escribir codigo. Tambien usar para pedidos como "haz un plan", "planifica este feature", "scope this implementation" o "before coding".