Mis Reglas Para Que Claude Code No Me Mienta

Value of the Information & Strength of the Argument

The video offers valuable, actionable insights for developers using AI coding assistants. The hosts present a clear taxonomy of hallucinations and back it with a referenced study, enhancing credibility. They demonstrate real-world defenses, such as MCP integration and Git hooks, which are practical and immediately applicable. The argumentation is persuasive, grounded in personal experience and specific examples, though it lacks rigorous empirical validation. The discussion of the ‘improve’ skill with subagents is innovative and thought-provoking, though its effectiveness is not empirically proven.

Scientific Rigor, Source Quality, Title Accuracy

The video references the CodeHallu study from the University of Hong Kong, providing a scientific basis for the taxonomy. However, the study itself is not directly cited with a URL, and the hosts rely heavily on anecdotal evidence. The description includes links to tools like skills.sh, ztrace, and claude-cron, which are relevant and useful. The title accurately reflects the content, focusing on rules to prevent Claude Code from lying. The video’s scientific rigor is moderate: it presents a structured framework but lacks formal citations for many claims. The hosts do not provide a systematic evaluation of the techniques’ effectiveness, but they do offer practical guidance.

The title accurately reflects the content, which focuses on rules and techniques to prevent Claude Code from hallucinating.

7/10

The video provides practical, experience-based advice on mitigating LLM hallucinations in coding, referencing a specific study (CodeHallu) and demonstrating real tools. However, claims are largely anecdotal and not peer-reviewed, and some external links are not directly verified.

• 00:00Introducción
• 00:20El concilio de expertos: skill de panel de debate con IA
• 02:31Claude Code en 2026: por qué ya no usamos otra cosa
• 04:09Conectar Claude Code con Notebook LM (MCP)
• 08:52Demo en vivo: notebook de investigación sobre agentes IA
• 10:18Claude Code para cosas que no son código (Remotion, marketing, Meta Ads)
• 12:01CodeHallu: los 4 tipos de alucinaciones al programar
• 13:23Mapping: cuando cruza identificadores válidos
• 15:14Naming: se inventa nombres plausibles (solo 39% válidos en GPT-4o)
• 16:10Fantasma: campos inventados en APIs
• 18:20House of Cards: el castillo de naipes de errores acumulados
• 22:18Crear interfaces para máquinas, no forzar las nuestras (herramienta Z3)
• 23:04Git hooks: pre-commit con Ruff y pre-push con pytest
• 30:43Hooks de Claude Code: pillarlo con las manos en la masa
• 39:03/loop y /simplify de Anthropic + cron para Claude Code
• 43:05Skill "improve": 3 subagentes Jedi revisan tu código
• 48:03Skills.sh: el ecosistema de skills para todos los agentes
• 51:44Proyecto open source: administrador web para Ollama
• 53:07Reflexión final: no evitar alucinaciones, detectarlas a tiempo

• Skills.sh — Ecosystem of skills for AI agents, mentioned as a resource for extending Claude Code capabilities.
• Ztrace — Instrumentation tool for Mac, used to create machine-friendly interfaces.
• Claude Cron — Tool for scheduling Claude Code tasks, mentioned in the context of automation.
• Remotion — React-based video editing tool, used as an example of non-coding applications of Claude Code.
• Notebook LM MCP — MCP server for Notebook LM, enabling integration with Claude Code for delegated research.

• CodeHallu: A Study on Hallucinations in Code Generation — Referenced in the video as the source of the taxonomy of four hallucination types, though no direct URL is provided.