Code agents Claude Code and Codex can be guided by ‘skills’, small markdown files (plain text format) that instruct the agent on how to proceed in a given situation. Launched by Anthropic, these skills avoid having to constantly repeat the same architecture, convention, or methodology instructions. The JDN has selected four, to which it adds a fifth as a bonus.
Code Quality and Security
The Skills For Real Engineers library is among the most cited for the software development cycle. Its author, Matt Pocock, is a British developer and trainer, initially a global reference for TypeScript, Microsoft’s programming language. Since February 2026, he has made his 27 skills available on GitHub, designed based on the shortcomings he observed in Claude Code and Codex. Updated almost weekly, they guide the agent before complex development, limit verbose responses, and plan for regular, systematic testing of new code.
Among the skills most used by SaaS application developers (software accessible online by subscription), security-audit stands out. Published by Cloudflare in June and then updated several times, it analyzes a project’s architecture, searches for vulnerabilities by family, and checks if each is actually exploitable. It then produces a report with recommendations, where each result is classified as ‘confirmed’, ‘to be validated’, or ‘rejected’. It relies by default on sub-agents to dig deep into the code. While, according to the JDN, cybercriminals are massively using AI to spot vulnerabilities, this tool can help avoid deploying vulnerable code.
Memory and Plan Review
The Obsidian Skills plugin, developed by Steph Ango, the CEO of Obsidian, contains six skills. Obsidian is a markdown note-taking application that developers use to store all or part of their agent’s memory. The skills allow writing markdown in the Obsidian format, creating databases with views, filters, and formulas, designing visual tables, managing the vault via command line, or converting web pages to clean markdown. The plugin also reminds the agent of the correct format for each file, taking into account the latest evolutions of this format.
Published in October by a team of Google engineers, the Elephant-Goldfish Model method is based on an idea from Dave Rensin, a distinguished engineer at Google, one of the company’s highest technical ranks. A first agent, named ‘the elephant’, retains all the context and establishes the project plan. Several memory-less agents, the ‘goldfish’, then coldly review this plan, without having seen the conversation, to spot implicit assumptions, vague requirements, and implementation risks. The developer decides and validates before the agent writes the code, in order to correct design errors before discovering them in the code.
Anticipating Model Evolution
The future-proof skill is inspired by advice from Tibo Sottiaux of OpenAI: build your projects imagining that AI models will be about ten times better in a year. Before a new project or feature, a separate review agent evaluates it based on seven criteria, including the project’s ability to benefit from AI progress, its potential use by other agents, or its dependence on current prices. The skill assigns a score out of 10 for three, six, and twelve months. If the one-year score is too low, it proposes a revised plan.
