@shadouyoua: Recently, the ByteDance TRAE team released the '2026 Enterprise AI Programming Practice Manual', which includes a noteworthy section: their summarized 'Top 10 Agent Skills'. This is the first AI programming skill recommendation list I have seen that has been publicly compiled by a major tech company. ...
The ByteDance TRAE team has released the '2026 Enterprise AI Programming Practice Manual' and published an internally compiled list of the Top 10 recommended Agent Skills. This list highlights the importance of frontend design, code review, and automated testing, showcasing best practices from a major tech player in the field of AI-assisted programming.
Recently, the ByteDance TRAE team released the '2026 Enterprise AI Programming Practice Manual', which contains a particularly noteworthy section: their summarized 'Top 10 Agent Skills'. To the best of my knowledge, this is the first AI programming skill recommendation list publicly compiled and published by a major technology company.
Let's briefly set the context. TRAE is an enterprise-grade AI programming tool built by ByteDance on top of the Doubao-Seed-2.0-Code model. Its goal goes beyond simple 'code completion'; it aims to have AI genuinely participate in the entire software development lifecycle: breaking down requirements, task planning, coding, testing, and deployment, striving to function more like an 'AI Engineer'.
After extensive internal usage of Agents, they distilled the ten most frequently used Skills, ranked as follows:
1. frontend-design — Frontend Design
It comes as no surprise that this ranks first. While many AI-generated UIs may be functional, their aesthetics are often lacking. The core purpose of this Skill is to enable AI to produce more reasonable and visually appealing interface designs.
2. cache-components — Component Caching
This instructs the AI to prioritize reusing previously generated components rather than creating new ones from scratch every time. This significantly reduces token consumption and improves generation efficiency.
3. fullstack-developer — Full-Stack Development
This applies a 'full-stack engineer' mindset to the AI, encouraging it to look beyond single modules and consider frontend-backend collaboration, data flow, and API design simultaneously.
4. frontend-code-review — Frontend Code Review
5. code-reviewer — General Code Review
Two of the top five spots are held by review-related Skills. This indicates that ByteDance places high importance on the quality of AI-generated code internally—the focus is not just on whether the code 'runs', but whether it is 'reliable and maintainable'.
6. webapp-testing — Web Application Testing
7. pr-creator — Automated PR Creation
Automatically generates Pull Requests upon code completion, further automating the development workflow.
8. fix — Bug Fixing
Specifically designed for scenarios like 'There's an issue here, please fix it'.
9. update-docs — Documentation Updates
Ensures documentation is updated in sync with code changes, preventing discrepancies between docs and code.
10. find-skills — Skill Discovery and Search
This one is particularly interesting. It essentially functions as a 'meta-skill'—enabling the AI to identify missing Skills and recommend them autonomously.
After reviewing this list, several impressions stand out:
First, the fact that 'frontend-design' ranks number one suggests that many have been frustrated by the default AI aesthetic of 'purple gradients and rounded cards'. Anyone who has actually used AI for frontend development understands why this Skill is critical.
Second, the presence of two review Skills and one testing Skill in the Top 10 reveals ByteDance's clear philosophy regarding AI programming: coding speed is not the core metric; code quality is. This aligns closely with conclusions from the SkillsBench paper, which found that good Skills can improve Agent performance by 51%, whereas poor Skills can cause a 39% negative impact.
Third, 'find-skills' is the most noteworthy direction. It signifies that AI is evolving from merely 'using skills' to 'actively seeking skills'. This approach is very similar to the 'automatic skill evolution' concept proposed in EvoSkill—ultimately, the optimal Skill combination is likely not manually configured by humans, but evolved by the AI through long-term practice.
If you regularly use tools like Claude Code or Codex, this list is highly valuable for reference. Especially the core Skills at the top, which can essentially serve as a mature configuration strategy for Agents.
Original '2026 Enterprise AI Programming Practice Manual': https://lcnziv86vkx6.feishu.cn/wiki/XZOSwI51wi5a5okxCF4cAxHSnBh…
Shares 8 curated AI skills, covering basic configuration, product development, and content creation, to boost AI productivity for agents such as Claude Code and CodeX.
Introduces an open-source project called AGENTS Book Rules, which organizes the core principles of 13 classic programming books (such as Clean Code, Domain-Driven Design, etc.) into rule files that can be directly used with AI coding tools like Claude Code, Codex, and Cursor to improve code quality and architecture design capabilities.
The Agency is a GitHub repository with 232 specialized AI agent prompts and workflows for roles like frontend, backend, DevOps, and more, designed to help developers use AI coding tools more effectively.
ByteDance's VP of Technology Hong Dingkun shared reflections on AI Coding, pointing out that AI code contribution rate should not be a KPI, functional correctness does not equal engineering readiness, and the challenges of team collaboration after the lowering of the coding bar, emphasizing the importance of systematic AI development and foundational engineering.
A guide titled 'Agents Best Practices' providing a provider-neutral Agent Skill for building production-level AI Agent harnesses, designed for tools like Claude Code and Codex.