PromptKit: Reusable Agent Skills for AI-Assisted Engineering
How PromptKit evolved into a library of focused agent skills that make engineering workflows explicit, repeatable, and easier to review.
- #ai
- #developer-tools
- #architecture
- #react
Updated September 2026: PromptKit has evolved from a prompt library into a catalog of reusable agent skills. This article reflects the current project.
AI-assisted development is most useful when the task has a clear scope, the relevant context is available, and completion can be verified. Repeatedly explaining those expectations led me to build PromptKit.
Turn repeated guidance into a maintained artifact
PromptKit is a static library of agent skills for engineering work, including codebase analysis, implementation planning, debugging, code review, and .NET development. Each skill packages a focused workflow in a SKILL.md file that can be inspected, downloaded, and adapted.
The shift from individual prompts to skills reflects how I use AI: recurring work benefits from explicit activation criteria, ordered steps, boundaries, and a clear completion report.
Give each skill one responsibility
A useful skill should explain when it applies and what evidence it needs. A code-review workflow should direct attention to the changed behavior, surrounding contracts, plausible failure cases, and relevant verification. An implementation workflow should connect a concrete requirement to a reviewable change.
Combining every engineering activity into one instruction file makes applicability harder to judge. Focused skills are easier to revise when experience exposes an ambiguity or a missing check.
Keep the delivery simple
The current site uses React, TypeScript, and Vite to present a searchable catalog, skill details, and installation guidance. The skill files remain readable Markdown. A static site fits the job: people can inspect the instructions before deciding whether they belong in a personal or repository workflow.
The catalog includes guidance for several coding agents, with platform-specific installation details on the site. The underlying engineering expectations remain explicit in the skill itself.
Treat instructions as engineering work
A reusable skill can still be wrong, incomplete, or poorly matched to a repository. I evaluate it through the resulting changes: did the agent understand the relevant code, stay within scope, run meaningful checks, and report limitations accurately?
When a workflow produces a recurring mistake, the response is to investigate the assumption and refine the guidance. Adding more instructions without understanding the failure can make the next run harder to reason about.
Keep responsibility with the engineer
AI can help explore alternatives and execute well-defined work. I remain responsible for design decisions, reviewing the diff, and checking the behavior. Generated tests also need review: they can repeat the same mistaken assumption as the implementation.
PromptKit is one way I turn continued learning into a practical tool. The library captures workflows I can revisit, improve, and share as my experience with AI-assisted development grows.
Explore PromptKit or read my approach to code review.
