{"id":1768,"date":"2026-05-12T18:02:33","date_gmt":"2026-05-12T18:02:33","guid":{"rendered":"https:\/\/ftechpilot.com\/mis-skills-favoritas-de-claude-code-y-por-que-las-uso-cada-dia\/"},"modified":"2026-05-13T16:06:27","modified_gmt":"2026-05-13T16:06:27","slug":"mejores-skills-claude-code","status":"publish","type":"post","link":"https:\/\/ftechpilot.com\/en\/mejores-skills-claude-code\/","title":{"rendered":"My favorite Claude Code skills (and why I use them every day)"},"content":{"rendered":"

I have been using Claude Code almost daily for months and, over time, I have been filtering what skills<\/strong> They really change how I work and what is noise. This list of best extensions for Claude Code<\/strong> It is my personal opinion. No sponsorship, no affiliates, just what I actually use.<\/p>\n

I have tried them all in real projects: compressing prompts, humanizing text generated by AI, memories between sessions, reviewing code with several agents... each plugin contributes something specific. I have discarded dozens; These five have stayed.<\/p>\n

We talk about Caveman<\/strong>, Humanizer<\/strong>, Nadkin<\/strong>, Claude-mem<\/strong> y Ultra-review<\/strong>. Empezamos.<\/p>\n

\n

Caveman \u2014 when every token counts<\/h2>\n

The name is strange, I know. But what he does makes a lot of sense (JuliusBrussee\/caveman<\/a>): Compress your prompts so that Claude consumes as few tokens as possible without losing relevant information. In practice, it is as if you were writing in a telegram but Claude understood it perfectly.<\/p>\n

You notice it especially in large projects or long conversations where the context begins to become saturated. With Caveman, I can fit more things within the same limit. It's not magic, but it helps a lot.<\/p>\n

How to use<\/h3>\n
\/caveman\n\nYou are a wizard that helps optimize Python code for production. The project uses FastAPI, PostgreSQL and Redis. I have a bug on the authentication endpoint where JWT tokens are not invalidated correctly when logging out...<\/code><\/pre>\n
Claude compresses it himself before processing it. You write normally, he works with less.<\/p>\n
\n
Humanizer \u2014 because AI text is too noticeable<\/h2>\n
If you use Claude to write content, you already know the problem. That corporate PowerPoint tone, the sentences that start with \u201cIt is important to note that\u2026\u201d, the perfectly symmetrical three-point lists. Everyone recognizes it right now.<\/p>\n
Humanizer locates these patterns and eliminates them. It's not just removing filler words; rewrite to make it sound like someone real. I've tried other methods and none of them work as well as this one because it works with a very specific list of documented patterns (it's based on Wikipedia's guide on AI writing).<\/p>\n
This is the link to the repository (blader\/humanizer<\/a>). Judge for yourself.<\/p>\n
Quick example<\/h3>\n
\n\n\n\n\n\n\n
Before (pure AI)<\/th>\nAfter (Humanizer)<\/th>\n<\/tr>\n<\/thead>\n
\"It is essential to highlight that this solution represents a significant advance in the current technological landscape.\"<\/td>\n\"Works. And that's enough.\"<\/td>\n<\/tr>\n
\"It is worth mentioning that the results obtained underline the importance of adopting a holistic approach.\"<\/td>\n\u00abThe numbers improved. Probably because we stopped complicating it so much.<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n
\n
Nadkin \u2014 the AI \u200b\u200bthat learns from its own mistakes<\/h2>\n
This is the one that surprised me the most. When Claude makes a mistake, Nadkin records it in a markdown file. Not for that session. Forever. The next time you open Claude on that project, know what not to do.<\/p>\n
It seems small, but in long projects the experience changes a lot. Stop having to repeat \"I told you not to do this like that\" every two weeks. The accumulated context of the failures is saved and Claude consults it before acting.<\/p>\n
The file that generates<\/h3>\n
# Errores aprendidos \u2014 proyecto ftechpilot\n\n## CSS\n- No usar \\uXXXX en la propiedad `content` de CSS. El navegador lo interpreta\n  como texto literal. Usar el car\u00e1cter directamente: '+', '-', '>'\n\n## WordPress API\n- Los snippets run-once (11 y 12) NO se deben reactivar. Ya se ejecutaron.\n- Despu\u00e9s de cambiar CSS visual, avisar al usuario de hacer Purge All en LiteSpeed.\n\n## Snippets\n- El snippet de purga (9) debe quedar inactivo despu\u00e9s de usarse.\n  Activar \u2192 disparar \u2192 desactivar. Nunca dejar activo.<\/code><\/pre>\n
That's literally what it saves in my project.<\/p>\n
It does not have its own public repository \u2014 it is included in community skills collections such as anthropics\/skills<\/a>.<\/em><\/p>\n
Each line is an error that will not happen again.<\/p>\n
\n
Claude-mem \u2014 real memory between sessions<\/h2>\n
The classic problem with AI assistants: every time you open a new conversation, it starts from scratch (thedotmack\/claude-mem<\/a>). If you have a project with three weeks of history, you have to explain it again. That wastes tokens and time.<\/p>\n
Claude-mem saves relevant context between sessions in structured memory files. Who you are, how you prefer to work, what projects you have active, technical decisions made. Next time you open Claude, read that first.<\/p>\n
The most notable difference is not in the first session. It's in the tenth. When Claude already knows that you use strict TypeScript, that you don't want obvious comments in the code, and that this specific project has a particular architectural restriction. That accumulated context is worth a lot.<\/p>\n
Types of memory it saves<\/h3>\n