Agent skills live in a weird place, especially with `npx skills`. You find a skill in a registry, install it, invoke it, and blindly pull updates whenever the author pushes a change. That workflow makes sense for binary code or compiled libraries, but that’s not the case for skills.
When you treat a skill like an immutable package, you inherit all of the author's edge cases, bloated instruction sets, and arbitrary style choices. Because skills are ultimately just Markdown, treating them as black-box dependencies is holding you back.
It’s time to start owning your agent skills.
The Problem with Generic Skills
To make a skill popular, authors have to make it universal. They must consider different project setups as well as different harnesses and models. Claude Code, Codex, Github Copilot and OpenCode all behave differently.
Some popular skills, like superpowers, repeat their instructions in multiple forms (text, graphs, numbered list) to ensure that most harnesses and models will follow them most of the time.
What happens when you load that skill into your local workspace?
- Context Bloat: Your agent spends hundreds (or thousands) of tokens reading instructions for frameworks you don't use, databases you've never touched, and error states you'll never encounter.
- Instruction Dilution: Large Language Models follow instructions best when those instructions are concise and direct. The more generic edge cases a prompt contains, the higher the chance the agent strays from the core task.
- Unpredictable Updates: An upstream author tweaking a skill might subtly degrade how that same skill performs in your setup.
- Malicious Updates: Skills are a great place to inject malicious prompts. We have seen several npm supply chain attacks. NPM packages have more security in place than your skill sources.
The Secret: It’s Just Markdown
The moment you realize an agent skill is simply structured text wrapped in a .md file, the friction disappears. There is no build step. There is no underlying binary logic to break.
Taking ownership of a skill isn't "forking a complex codebase". It’s just editing a text file to better suit your needs.
When you find a skill that gets you 70% of the way there, don't leave it in a global installation directory. Install it to your project’s repository and make it yours. Treat it like a template or a boilerplate component, not a third-party module.
How to Adapt and Trim a Skill
Once a skill lives in your repository (under a .agent/skills/ directory for Codex or OpenCode or .claude/skills/ for Claude Code), strip away everything that isn't actively helping your project:
- Delete the Irrelevant Options: If a skill includes instructions for three different test runners, delete two of them. If it has instructions for tools you don’t use, remove them.
- Inject Your Project’s Standards: Replace general instructions ("use modern styling") with your exact requirements (if short) or reference files containing them.
- Hardcode Your Workflow: If a skill tells the agent to commit changes, but your team wants to review before committing, change that.
- Prune the Defensive Prompting: Community skills often spend half their length warning the model not to do bad things. If your specific stack or setup naturally prevents those issues, delete the warnings and keep the prompt lean.
From Consumers to Authors
Bringing skills in-house changes how your team works with AI. Instead of relying on generic public prompts, your skills become part of your tooling. They evolve directly alongside your software.
If you want to maintain company-wide skills to be used by multiple teams, keep in mind that they will likely adapt your skills to the project, if needed.
A lean 20-line Markdown skill written specifically for your repository will outperform a 300-line "universal" skill. This is especially true for “cheaper” models (Minimax, Deepseek, Gemini Flash Lite). These models excel in low costs or high speed, but may need more direct and clear guidance.
Stop waiting for skill maintainers to fix issues that are unique to your setup. Pull the Markdown into your project, open your editor, and start adapting your skills to your project.
Do you need help taking ownership of your AI tooling and workflows?
Let's talk about how we at Intenics can empower your team without risk or dependencies.
