跪拜 Guibai
← All articles
Frontend · Coding Standards · Architecture

Agent-Skills Slows Down AI Coding by Encoding Senior Engineers' Workflow Habits

By 东方小月 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

As AI coding tools become default in daily development, the bottleneck shifts from code generation speed to engineering judgement. Agent-skills provides a lightweight, tool-agnostic way to inject process discipline into AI-assisted workflows, reducing the downstream cost of overconfident, under-scoped implementations.

Summary

AI coding assistants default to writing code immediately, often before requirements are clear. Agent-skills counters that by packaging engineering habits into invocable skills: spec-driven development, planning and task breakdown, test-driven development, code review, debugging with root-cause analysis, frontend engineering, performance optimization, and launch checklists. Each skill constrains the AI to follow a structured workflow — clarify requirements, break down work, verify before implementing, review for edge cases — rather than racing to produce code. The skills install via npx or plugin marketplaces for tools like Codex, and are invoked inline with @-mentions during a session. The result is a slower but more reliable development cadence that surfaces ambiguity early and catches risks before they become bugs. The project targets independent developers, small teams without formal process, and heavy AI-coding users who have learned that prompt brevity is less valuable than workflow discipline. The core insight is that AI programming problems cannot be solved by stronger models alone; workflow constraints are equally critical, much as junior engineers need process guidance more than raw coding ability.

Takeaways
Agent-skills is a collection of workflow patterns that constrain AI assistants to follow engineering process before writing code.
Skills include spec-driven development, task breakdown, TDD, code review, debugging with root-cause analysis, frontend engineering, performance optimization, and launch checklists.
Installation works via `npx skills add addyosmani/agent-skills` or through plugin marketplaces for Codex and similar tools.
Each skill is invoked inline with an @-mention, e.g., `@spec-driven-development` to force requirement clarification before implementation.
Using spec-first and review workflows surfaces ambiguity and edge cases early, which prevents rework on complex features.
The debugging skill guides AI to reproduce, narrow scope, check logs, and find root causes instead of patching surface symptoms.
Code review via AI often delivers more value than code generation because it works from existing code rather than inventing from scratch.
The project is most useful for independent developers, small teams without established process, and heavy daily users of AI coding tools.
Conclusions

Stronger models alone won't fix AI coding quality; without process constraints, a powerful model can still produce an unmaintainable mess, while a weaker model inside a good workflow can deliver stable value.

The parallel between junior engineers and AI assistants is structural: both can write code, but neither knows when to stop and question requirements, check for ambiguity, or verify assumptions before acting.

Code review is an under-exploited AI use case — it requires analysis of existing code rather than creative generation, which plays to current model strengths while avoiding hallucination-prone synthesis.

The project's value proposition is explicitly about slowing down, which runs counter to the industry's default framing of AI as a speed multiplier, and that inversion is what makes it worth attention.

Concepts & terms
Spec-driven development
A workflow where functional specifications — covering types, actors, edge cases, and constraints — are written and agreed upon before any implementation code is produced.
Agent skills
Reusable, invocable workflow patterns that constrain an AI coding assistant's behavior to follow a specific engineering process, such as TDD or code review, rather than generating code freely.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗