Estás usando Claude Code mal: así lo usan los ingenieros de Anthropic
This video explains how Anthropic's own engineers use Claude Code through a skills-based system rather than isolated prompts. The core idea is to build reusable, modular skill folders that package instructions, examples, and scripts for repeated tasks. Four principles guide this approach: create skills for repeated tasks, save complete processes not just instructions, keep skills small and specific, and continuously improve skills based on corrections.
Summary
The video presents four principles derived from how Anthropic's engineers actually use Claude Code, contrasting this approach with the common habit of writing new prompts from scratch each time.
The first principle establishes that repeated tasks should not depend on new prompts each time. Anthropic created 'skills'—organized folders that package procedural knowledge—so Claude doesn't start from scratch. A skill can contain instructions, examples, reference files, documentation, rules, and executable scripts. The presenter illustrates this with a visual presentation skill that encodes the entire workflow, so instead of writing a long prompt each time, the user simply says 'turn this document into a presentation.'
The second principle is that a skill should contain the entire process, not just a single long instruction. Anthropic's Skill Creator tool can help build skills, but a good skill has three layers: a description (telling Claude when to activate the skill), instructions (what steps to follow once activated), and resources or tools (reference files, templates, scripts). The presenter cites an Anthropic example where engineers noticed Claude repeatedly writing the same Python script to style slides, so they saved it inside the skill to avoid redundant regeneration.
The third principle warns against creating one massive skill that tries to do everything. Instead, skills should be small, specific, and composable. A 'content creator' mega-skill is hard to debug and maintain, whereas separate skills for researching ideas, analyzing transcripts, writing scripts, and creating hooks each have a clear purpose. This modularity means improvements to one skill benefit all contexts where it's used, and skills can be chained together as needed.
The fourth principle is iterative improvement: after every use of a skill, the user should ask whether corrections made in the chat should be permanently incorporated into the skill itself. Anthropic's stated goal is for Claude after 30 days of use to be significantly better than on day one. The presenter recommends asking Claude to review recent conversations and identify which repeated corrections should become permanent rules in the skill, turning the system into one that compounds in quality over time.
Key Insights
- Anthropic engineers explicitly stopped building increasingly complex agents and shifted to building 'skills' instead—organized folders that package composable procedural knowledge for Claude to reuse across tasks.
- Anthropic observed that Claude was repeatedly generating the same Python script to style slides across conversations, so engineers saved the script inside the skill itself so future Claude instances could simply run it rather than regenerate it.
- A skill's description layer is what Claude checks to determine when to activate the skill, and if written well enough, Claude can auto-activate the skill without the user explicitly invoking it—for example, detecting that 'convert this document into slides' warrants the presentation skill.
- Anthropic's documentation emphasizes that skills should not be overloaded with context; a mega-skill is harder to debug because when something fails, it's unclear which layer caused the problem, whereas small specific skills isolate failures and allow targeted improvements.
- Anthropic frames the goal of the skills system as Claude after 30 days of working with a user being significantly better than Claude on day one—not because Claude learns autonomously, but because corrections are systematically written back into skills as permanent rules.
Topics
Full transcript available for MurmurCast members
Sign Up to Access