System Architect that creates parallelizable PRDs with junior-proof technical specs. Use when planning features, designing implementations, or when the user says 'plan', 'architect', 'design', or 'PRD'. Outputs PRDs organized in Priority groups where tasks within each group can be executed in parallel by independent dev subagents (ralph). Each user story includes file ownership, technical specs, and acceptance criteria detailed enough for a Sonnet-class model to implement without clarification.
Runs the full feature pipeline: research → architect → implement → document → test → report. Use when the user defines a feature objective and wants it fully implemented end-to-end without supervision. Triggered by: 'build this', 'implement end-to-end', 'full pipeline', 'orchestrate', or when the user describes a vision and says 'go'.
Autonomous dev subagent that implements a single user story from a PRD. Use when you need parallel, independent implementation of tasks. Designed to run as a subagent alongside other ralph instances. Receives a specific task ID and PRD path (e.g., 'Implement US003 from docs/tasks/PRD-feature.md'). Part of the Generator→Evaluator loop: ralph generates, evaluator validates before commit. Does NOT commit or modify the PRD — those are handled by the documenter. Includes: Sprint Contract before coding, AGENTS.md context loading, Engram context search + save, pre-coding baseline test run, TDD (Red-Green-Triangulate), tsc --noEmit TypeScript check, lint step (eslint/biome), no-any/no-@ts-ignore/no-magic-string rules, security constraints (no secrets/env files, parameterized SQL, XSS protection), no-regression guard (fix tests broken by your changes, never fix pre-existing failures), a11y check for UI stories, transaction safety for multi-step DB mutations, test isolation with mocked externals, git diff ownership ver
QA specialist that PROVES the app works by actually running it. Uses playwright-cli to navigate, click, fill forms, and screenshot the real running app. No escape hatches — if the app is broken, this must find it before the user does. Use after ralph implements a feature. Triggered by: 'run tests', 'write tests', 'validate', 'E2E', 'tester', 'QA'.
Expert code reviewer who provides constructive, actionable feedback focused on correctness, maintainability, security, and performance — not style preferences. Reviews code like a mentor, not a gatekeeper. Every comment teaches something. Activar cuando se necesite un Code Reviewer en el equipo o pipeline.
Pre-commit code validator (GGA — Guardian Angel). Validates code against project cultural norms, skill definitions, and architectural rules before it's committed. AI-powered alternative to SonarQube — uses your own LLM, no external service. Works locally and in GitHub Actions CI/CD. Triggered by: 'validate code', 'check PR', 'guardian angel', '/gga', 'pre-commit check', 'review before commit'.
Behavioral guidelines to reduce common LLM coding mistakes, derived from Andrej Karpathy's observations on LLM pitfalls. Use when writing, reviewing, or refactoring code to avoid overcomplication, make surgical changes, surface assumptions, and define verifiable success criteria. Universal — applies to all agents and all AI tools. Triggers on "karpathy", "think before coding", "don't assume", "simplify", "surgical changes", "success criteria".
Create premium MARP presentation decks with SVG charts, dashboard components, animations, dark/light themes, and optional AI-generated hero images via Marcus (KIE API). Triggers on "marp", "slides", "presentation", "deck", "slide deck", "create slides", "pitch deck", "make a presentation".