Why I Built an AI Twin to Run My Life

Last reviewed: 5 November 2026.

“AI will save you so much time!” Silicon Valley promised. So I tried everything. ChatGPT for writing assistance. Claude for research summaries. Gemini for project management. Copilot for planning. Every morning started the same way: ten minutes explaining my context to whichever AI I was using that day.

“I’m Director of Product and Innovation, Sub-Saharan Africa, at Cambridge University Press & Assessment…”

“I need this in British English for an education audience…”

“Keep it practical, I’m building training programmes…”

Then I’d finally get to the actual work. AI was supposed to save time. Instead, I felt busier than ever.

I was drowning in AI tool overload while still doing everything manually.

The working laboratory is now NiallOS. The method behind it is expertise encoding. The ownership argument that came out of this work is Own the Memory, Rent the Intelligence.

The AI promise vs reality

In 2023 and 2024 the pitch was simple: AI will transform knowledge work, automate everything, and free up your week. My reality in a Cambridge product and innovation role looked different. I was designing training programmes for education leaders, working with organisations such as UNICEF and KFAS, creating workshop materials and curricula, publishing public writing on AI in education, and staying current on EdTech practice across Sub-Saharan Africa.

AI was supposed to save time on routine tasks, help me focus on strategic work, make complex work manageable, and give me breathing room.

What actually happened? Every morning became an AI context-switching marathon. ChatGPT for brainstorming workshop activities. Claude for analysing research papers. Back to ChatGPT for LinkedIn drafts. Gemini for curriculum outline suggestions. Each tool required a fresh start and a fresh explanation of who I was and what I was building.

I calculated the daily context tax. I wasn’t using AI to work smarter. I was managing a portfolio of AI tools, each one needing setup, context and translation. It felt like having five junior assistants who all had amnesia every morning.

Claude Code terminal open while building a personal AI operating system
Claude Code becoming the shell for a personal operating system

The breakthrough

I was watching a podcast where a startup founder demonstrated how he’d set up his Claude Code life management system. He showed slash commands originally designed for software development. /commit generated git commits. /review-pr analysed code changes. Custom commands ran his project patterns.

What caught my attention was how he had adapted those development tools for daily work. No re-explaining project structure. No describing preferences. The system already knew his contexts, conventions and workflow.

That is when it clicked: what if I could build slash commands for my actual job?

Claude Code offered three things generic chat tools did not.

First, access to the file system. It can read project files, see folder structure, and pick up context from existing documents. No manual copying and pasting.

Second, access to the terminal. It can run commands, execute workflows and integrate with real tools. That is automation, not only text generation.

Third, persistent customisation. You can create slash commands, encode preferences, and let the system evolve with how you work.

Generic chat tools treat you like a stranger every day. A local, file-aware setup can read your work, keep your conventions and run your workflows the way you actually work. That was a foundation for a personalised AI operating system.

Building NiallOS

I spent a weekend asking: if I could teach AI everything about how I work, what would I tell it? Not one-off instructions. A foundation profile every workflow could read first.

What I built is NiallOS, my personal AI operating system. It is not an app and not a productivity hack. It is folders and markdown files on my machine, with constant backup, so I could point another model or CLI at the same structure if I needed to.

List of custom slash commands used inside NiallOS
Some of the custom slash commands

The digital twin

I created a profile file that every AI workflow reads before it runs.

Professional identity covers my Cambridge role and focus on AI programmes for education leaders across Sub-Saharan Africa. Expertise spans EdTech, AI literacy and learning design. Current projects at the time included UNICEF Innovation Camp work, the KFAS Future Educator programme and Cambridge digital leadership programmes.

Writing standards specify British English, plain explanation for practitioners, and careful citation when I draw on published research. Communication preferences are clear: practical, accessible, enough detail without padding. The primary audience is education leaders, practitioners, ministries, funders and partners.

Work patterns matter too. Peak focus sits early in the morning. Context switching is expensive. Programme delivery and partner commitments come first. Public writing and professional development sit around that core.

That twin holds professional context, standards, constraints and goals. Once it is in place, every workflow starts there. No more re-explaining. No more context tax.

Slash commands

With the twin in place, I built custom workflows as slash commands.

Daily workflows include a morning brief on EdTech, AI in education and policy, and an evening check-in for reflection and open loops.

Weekly systems track progress across projects and keep a simple dashboard of active work.

Research and writing commands organise sources, draft in British English for a named audience, and keep citations honest when a claim needs them.

Learning commands build short skill paths and review what stuck. Content and training commands plan workshops, shape curricula and turn session notes into portfolio writing.

Specialised agents

Each slash command launches a focused agent that reads the twin first, then does one job: progress analysis, source organisation, workshop design, learning-path design, or adapting content for a different format. The point is the same throughout. Instead of adapting myself to generic AI tools, I built a system that adapts to me.

Why personalisation changes the work

Every generic AI conversation starts fresh. “I need help planning a workshop on AI literacy…” Then come the questions: audience, level, length, goal. Five minutes of context before any useful help.

With a personalised system I type something like /workshop-planner ai-literacy and the output already knows the audience is education leaders and policymakers, the format is instruction then demo then practice, the style is practical and story-driven, and the usual length is 90 minutes. Zero context-setting. Straight to a draft I can judge.

Setup costs a few hours for the twin and the first commands, then ongoing refinement. Returns show up every time a command skips 15 to 45 minutes of re-explanation, keeps quality aligned with your standards, and stays consistent because the same framework runs each time.

Who benefits most? People who repeat the same kinds of work with clear preferences: educators designing training, programme managers juggling projects, researchers organising literature, anyone who regularly re-explains context to AI. Generic tools ask what you want. Personalised systems already know who you are, what you are working toward, and how you work.

Where this went next

This post was the origin story. The system has since been simplified and rebuilt more than once. What stayed useful was not a long list of slash commands. It was the idea that memory, method and model should be owned deliberately rather than rented by accident.

If you want the current shape of the laboratory, start with NiallOS. For the method of capturing how an expert actually works, see expertise encoding. For the three-layer ownership test that came out of this experiment, read Own the Memory, Rent the Intelligence.

This is not about copying my folders. It is about teaching AI your context once, building workflows that adapt to you, and treating memory as infrastructure rather than as another chat window.

Niall McNulty is Director of Product and Innovation, Sub-Saharan Africa, at Cambridge University Press & Assessment.


Discover more from Niall McNulty

Subscribe to get the latest posts sent to your email.

Leave a comment