NickOS: My AI Operating System

Overview

Most people probably think better AI output comes from better prompts. There is some truth in that, but I increasingly think context matters more: whether the AI already understands your work, your terminology, your stakeholders, your preferred formats, your tone, and the standards you apply. A good prompt can improve one output, but a good context system improves every output after that.

I call mine NickOS. It is not an app, but one folder containing many sub-folders and lightweight Markdown (.md) files, all in plain text. Markdown works well because it gives AI tools clean structure through headings, lists and links, without the formatting noise of PDFs, Word documents or slides.

At the root of the NickOS folder, I keep instruction files for the tools I use: CLAUDE.md for Claude, CODEBUDDY.md for CodeBuddy, and AGENTS.md for Codex. Each file briefs the relevant tool on how to work with the system. It also acts as a router, pointing the tool to a specific skill when I use a slash command such as /industry-tracker, or when a known term appears in my prompt.

The rest of the system is split into a few simple layers. MEMORY.md tracks active projects and anything I expressly ask the system to remember, usually transient notes tied to live work. _context/ stores my writing style, preferred voice, glossary and stakeholder profiles. skills/ contains one folder per repeatable task. projects/ contains one workspace per active project, each with its own instructions, source materials and project memory.

The folder structure is not the point. The point is that the AI no longer starts cold. It already knows how I write, which terms I prefer, who the output is for, what format I usually use, and what constraints matter. The same prompt, put to the same model, comes back sharper once the model has that surrounding context.

My Use Cases

Use Case 1: Translation

As an in-house counsel at a Chinese technology company offering products and services outside China, I work between English and Chinese on a daily basis. My Chinese is functional, but nowhere near native fluency. More importantly, legal and gaming terminology cannot always be translated literally. A term may be technically correct, but still sound unnatural to the stakeholders who actually need to use it.

To bridge that gap, I built a translation skill. It includes a gaming-industry glossary, preferred internal terminology, and examples of well-written Chinese industry updates, so that the output reflects a more natural professional register. It translates incoming Chinese requests into English while preserving legal and gaming-specific meaning, and converts my English drafts into Chinese that reads more naturally for internal stakeholders.

Here is a simple example of the translation skill in use. I applied the same prompt in two setups: Claude’s incognito chat, which does not use existing memory and relies on the base model alone; and Claude Code, with the NickOS folder loaded.

The incognito translation was broadly correct. It rendered “loot box” as terms such as 抽奖盒 or 战利品箱, which are common online translations and not necessarily wrong. Internally, however, the preferred term is 宝箱.

“Product team” is similar. The literal translation is 产品团队, but in my working context, 项目组 is often more natural because product team and project team often overlap in how colleagues refer to the relevant group.

The difference is not the prompt. The prompt is identical. The difference is that NickOS carries the glossary, context and examples into the task.

The base model gives a generally correct translation. The system gives a translation that is closer to immediately usable.

Use Case 2: Legal Brief

I also built a legal-brief skill that turns legal or industry developments into structured internal briefs.

When something relevant to the gaming industry catches my eye, I point the skill at the source document or link. It returns an executive summary, key points, gaming-industry implications, and recommended next steps.

For example, I tested it on the landmark US social media addiction verdict in K.G.M. v. Meta et al., where a Los Angeles jury found Meta and YouTube liable in relation to alleged addictive product design and failure to warn. I pointed Claude at a report on the case using my legal-brief skill, and the brief below came back. It gave me a quick read on the issue, and a solid first draft if I later decide to turn it into a polished stakeholder update.

Other Use Cases

Translation and legal briefs are only two skills among many. I also have a design skill that produces HTML and Word files in the formatting and aesthetics I prefer; a debrief skill that captures what is worth remembering after a session with any AI tool; and an audit skill that periodically prunes memory and instruction files so NickOS does not bloat.

None of this is finished, and that is the point. I keep creating new skills inspired by what others share on GitHub, what colleagues build internally, and what my own use cases reveal. Just as importantly, I keep pruning what is no longer useful and tweaking NickOS to better understand how I think and work. Over time, it becomes less a set of isolated AI interactions, and more a context system that keeps adapting to how I work.