Artificial Intelligence

From Writing to Doing: 10 GitHub Agent Skills That Make AI Code for Real

By 雨中飘荡的记忆 · Oct 26, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Agent Skills turn AI assistants from code generators into workers that follow documented procedures, which cuts the regression risk of unchecked AI edits and makes the assistant's output auditable in a real project.

Summary

AI coding assistants still stumble on real projects: they skip security checks, expand fixes into rewrites, and ship generic pages. Agent Skills are a response: structured "job manuals" that lay out steps, constraints, and verification methods for a class of task, so an assistant doesn't just answer but completes the work reliably.

This roundup, based on a Skill Leaderboard from October 25, 2026, tracks GitHub star gains over seven days, not absolute quality. The ten repos span security auditing, engineering habits, code knowledge graphs, technical illustration, front-end design, diagram generation, and a disciplined development process. Cloudflare's security-audit-skill, for example, forces assistants to check auth, input validation, and sensitive data handling before merge. Ponytail keeps AI from over-refactoring, while Superpowers enforces clarify-design-plan-test discipline.

The rest focus on polish: impeccable and taste-skill refine AI-generated interfaces; anatomy and diagram-design turn concepts into visual assets; graphify maps a codebase's relationships. Lemo-opuscar experiments with code-driven storytelling. Each repo serves a different phase of work, and the pattern is the same: move AI from generating code to following repeatable, verifiable procedures.

Takeaways

— cloudflare/security-audit-skill runs security checklists on auth, input validation, sensitive data, file uploads, and dependencies before merge.

— DietrichGebert/ponytail forces AI to explain reasoning and call chains before making minimal changes, preventing bug fixes from ballooning into refactors.

— mattpocock/skills collects reusable engineering workflows for requirements, testing, and code review instead of ad-hoc long prompts.

— Graphify-Labs/graphify maps code, concepts, and dependencies into a graph to help navigate unfamiliar large repos.

— wheresryan22/anatomy generates interactive technical illustrations for blogs and team explanations, with a warning to verify accuracy before publishing.

— pbakaus/impeccable reviews AI-generated front ends on layout, hierarchy, copy, and responsive behavior, and suggests changing one issue type at a time.

— lemomo-ai/lemo-opuscar uses code and Agent workflows to organize cinematic content, pacing, and effects, with dependencies to check in README.

— Leonxlnx/taste-skill refines fonts, spacing, color, consistency, and motion for AI pages that have functionality but lack visual polish.

— cathrynlavery/diagram-design picks the right chart type (flow, architecture, timeline, relationship) and requires diagram accuracy against the real system.

— obra/superpowers enforces clarify-design-plan-test discipline, checking results after each phase rather than generating everything at once.

Conclusions

The prominence of security and scope-limiting skills suggests the pressing problem with AI coding is not capability but control: developers need guards that stop AI from skipping checks or expanding changes.

The cluster of design, diagram, and illustration skills points to a second wave where AI output is judged on presentation and clarity, not just function.

Treating skills as version-controlled, shareable artifacts moves AI procedure from prompt folklore to documented engineering, similar to how libraries standardize code reuse.

Concepts & terms

Agent Skill

A structured set of instructions, constraints, and verification steps that an AI agent follows to complete a specific class of task reliably, as opposed to a one-off prompt.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗