02 Aug 2026
Adapting AI Skills to My Own Workflow
Overview
Most skills in NickOS, my AI operating system, begin with a task I find myself asking AI to help with again and again. Once the repetition becomes obvious, I codify the process into a skill so that I can call on the same workflow without explaining it from scratch each time.
Others begin with ideas I encounter elsewhere, often through a well-regarded GitHub project or an AI creator I have followed for a long time. I take the parts that fit how I work, adapt them to NickOS and leave the rest.
My /design skill followed both routes. It began with my own use case: I needed a consistent visual language across the HTML pages, dashboards, slide decks and other deliverables I kept generating. I used to design webpages, so I already had a sense of what I liked. The frustrating part was having to translate that taste into the same corrections each time. AI could produce functional work, but the visuals often slipped into excessive cards, weak hierarchy and generic software landing-page layouts.
From DESIGN.md to /design
Once I knew what I wanted the skill to solve, I fed the agent several DESIGN.md files I liked and asked it to rebuild or adapt them for my outputs. Those files did most of the initial heavy lifting by spelling out typography, colour, spacing and components; my role was to choose the references and decide what felt right.
That process drew on the wider DESIGN.md approach: a reusable Markdown brief that gives an AI coding agent a defined visual language. Sites such as getdesign.md and designmd.ai collect design systems that can be studied or applied to AI-built projects.
My use case was broader than applying one design system to one website. I wanted /design to choose an appropriate visual language for a legal brief, interactive form, dashboard, poster or cinematic deck. I gradually created variants around designs I kept returning to: Editorial Dark for cinematic pages and decks, Swiss Ink Light for analytical briefs, and Indigo Precision Light for forms and tools. More specialised variants cover posters, image-led pages and long-form narratives.
The skill now applies those variants through four modes: building an HTML deck, creating a new page or tool, polishing an existing page, or converting a Word or PowerPoint source into designed HTML.
Adapt selectively
Once /design became useful, I treated other people’s skills as sources of spare parts rather than packages I had to install. Paul Bakaus’s Impeccable helped sharpen the critique, audit and polish checks. Leonxlnx’s Taste Skill contributed ideas about density and layout variation. I left out overlapping commands, image-generation tools and motion-heavy features that did not suit my work.
Those additions sit on top of the design system I already use. /design still makes the core choices about visual variant, page shape, density and whether the output needs any illustration or motion. It then loads only the relevant reference files. I get the benefit of external ideas without creating several competing design commands inside NickOS.
The NickOS dashboard and my HTML slide decks show how differently the same skill can behave. The dashboard is an interactive dark interface built around the Editorial Dark variant.
At the other end, Replacing PowerPoint Decks with HTML Files uses a light editorial variant for a very different output. Its title card is rendered below as live HTML rather than a screenshot.
What I like most about my bespoke design/ skill is that it gives my visual preferences a reusable form, so the agent can start much closer to what I like while I still make the final judgement.
Other Skills, Same Pattern
The habit has spread beyond design, and the source does not need to begin as a skill.
Fable Mode. I adapted a five-gate method shared by AI automation creator Nate Herk in How I Make Opus Think Like Fable. It makes difficult work pass through scope, evidence, adversarial reasoning, verification and reporting. I connected it to existing NickOS checks and kept it behind an explicit command so simple work stays simple.
Legal STORM. Stanford OVAL’s STORM uses perspective-guided questions and simulated conversations to produce cited research. I adapted that architecture so several agents examine a difficult legal question from different stakeholder positions, while separate agents verify the important authorities. NickOS adds the legal evidence hierarchy, confidentiality controls and citation checks.
One of the more enjoyable parts of building NickOS is the discovery. Every so often, I browse GitHub Trending or watch an AI builder I follow on YouTube or Instagram, and come across a workflow I had not thought to look for. Fable Mode arrived that way: through a video and community download rather than a public repository.
When something catches my eye, I ask my agent to explain the method and separate the reusable idea from assumptions tied to the original author’s tools or audience. Most ideas remain references. The useful ones are folded into an existing skill or become the starting point for a new one.
This gets easier once the operating system already holds my writing style, terminology, project knowledge and general rules. I am not building a new assistant each time. I am teaching the existing one one more repeatable way to work.
Back to AI Journal