The latest articles on DEV Community by Debabrata Saha (@debabrata_saha). https://dev.to/debabrata_saha https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4089521%2Fb87a9701-542a-4afc-b8e4-2a404df8b2f3.png https://dev.to/debabrata_saha en Debabrata Saha Sat, 22 Aug 2026 10:20:30 +0000 https://dev.to/debabrata_saha/making-ai-coding-agents-actually-good-at-software-engineering-fi3 https://dev.to/debabrata_saha/making-ai-coding-agents-actually-good-at-software-engineering-fi3 <p>AI coding agents generate decent code. But decent ≠ production-ready.</p> <p>They miss the stuff experienced engineers know instinctively — accessibility standards, proper auth flows, database reliability patterns, API design conventions. The gap between "it works" and "it's production-grade" is real.</p> <h2> The Idea </h2> <p>What if we could give AI agents that missing knowledge as <strong>loadable skill modules</strong>?</p> <p>That's what <a href="https://github.com/debabratasaha-dev/techskills" rel="noopener noreferrer">TechSkills</a> does — an open-source library of <code>SKILL.md</code> files containing structured workflows, design checklists, and battle-tested patterns that any AI agent can consume.</p> <h2> How It Works </h2> <p>Skills use progressive loading to stay context-efficient:</p> <ol> <li> <strong>Metadata</strong> — ~100 words, always visible, triggers activation</li> <li> <strong>Core instructions</strong> — loaded on activation, kept lean (&lt;500 lines)</li> <li> <strong>References</strong> — deep checklists and patterns, loaded only when needed</li> </ol> <p>No context window waste. Agent loads what it needs, when it needs it.</p> <h2> Current Skills </h2> <ul> <li>🎨 <strong>frontend-engineer</strong> — responsive layouts, accessibility, design systems, React/Vue/Svelte</li> <li>⚙️ <strong>backend-engineer</strong> — APIs, auth, databases, reliability, observability, security</li> </ul> <h2> What Makes It Different </h2> <ul> <li> <strong>Agent-agnostic</strong> — works with any AI tool that reads markdown</li> <li> <strong>Framework-neutral</strong> — pseudocode patterns, not framework-specific recipes</li> <li> <strong>Eval-driven</strong> — every skill includes eval cases to verify it actually works</li> </ul> <h2> Early Stage, Looking for Contributors </h2> <p>This is Day 1. Two foundational skills. The vision is a community-curated library covering DevOps, system design, data engineering, mobile, and more.</p> <p>If you've got domain expertise and want to turn it into a skill module others can use — contributions are very welcome.</p> <p>⭐ <a href="https://github.com/debabratasaha-dev/techskills" rel="noopener noreferrer">GitHub</a> | 📄 MIT Licensed</p> <p>What skills would you want your AI agent to have? Drop ideas in the comments 👇</p> ai opensource productivity programming