What This Is
A developer log published on Juejin this week describes a painfully specific problem: the author had installed so many "Skills" (capability extension plugins for AI coding assistants — think of them as apps for your AI) that their description text was consuming so much of the AI's context window that the model's actual performance degraded. More plugins, d umber AI.
His fix: he designed a custom
file management structure and built a command-line tool called
hk-skill that handles the full plugin lifecycle — fet
ching, security review, adaptation, registration, and per-project activation. The whole thing is
open-sourced. What makes it worth noting is
that the tool itself was built using multiple AI models working in concert. He
documented in detail how he divided responsibilities across GPT, Kimi, and
GLM — planning, execution, and review — and what he learned from each.
The Industry View
This story hits on a genuine structural problem in the AI tooling world: plugins and tools are proliferating rapidly, but the infrastructure for managing those tools is essentially nonexistent. The plugin ecosystems of leading AI coding assistants — Cursor, Claude Code, and others — are still in a "wild growth" phase: no unified version control, no security review standards, no cross-project activation management. The pain point this developer ran into is one that any team using AI coding tools heavily will eventually hit.
That said, we should be clear about the article 's limitations. This is a personal project dev log. The tool's stability and maintainability are unproven, and the author himself admits he " started treating it as just a small utility," a mindset that led to a sharp spike in AI token consumption later on due to missing structured documentation. More fundamentally, his solution — a local file structure with symbolic links — is a developer-native, hands-on approach that offers virtually no usability for anyone without a technical background. History shows that "developer self -rescue tools" like this fail to become general-purpose products far more often than they succeed.
There 's a second observation in the article worth pulling out separately: his account of multi-model collaboration. Using GPT for planning and review, and Kimi for execution, outperformed using any single model alone. This remains a personal perception without quantitative backing, but it aligns with a trend we 've been tracking — experienced users are increasingly exploiting the task -specific performance differences between models deliberately and systematically.
Impact on Regular People
For enterprise IT teams: If your team is already using AI coding tools at any meaningful scale, "plugin and prompt version management" needs to go on your technical debt list now. This is not a minor housekeeping issue — it directly affects the stability and reproducibility of AI-assisted development workflows .
For individual professionals: The most practically useful part of this article isn't the tool itself — it's the author's hard -won lesson: "Design documents must come paired with a technical plan. Skipping that upfront costs far more later." This holds equally in any AI-assisted workflow. The more you cut corners on upfront structure , the higher the rework cost down the line.
For the broader market: The chaos in AI plugin management will inevitably attract commercial products to fill the gap. Whoever ships a credible "app store plus package manager for AI tools " first has a real shot at owning a foundational layer of the next-generation developer stack — and right now, that position has no clear frontrunner.