In my previous articles, I looked at two big ideas:

1. Why context is critical for getting high-quality outputs from AI

2. Why AI needs memory to enrich that context

Those ideas set the stage for the experiment I’m starting now: externalising my own memory so it can eventually be used by AI when solving tasks.

Starting With My Own Notes

I want to build what’s often called a Second Brain — an external memory I can rely on. For me, that means taking my messy collection of personal notes and slowly transforming them into a structured, AI-readable knowledge base.

Right now, I’ve got more than 1,000 notes sitting in Apple Notes.

It’s not a system — it’s a data lake. A place where thoughts, quotes, market insights, and half-finished ideas live side by side, waiting to be organised.

And this “lake” is where my journey begins.

The Roadmap

When I thought about turning my personal notes into something closer to an external memory, I knew I needed a roadmap.

For me, it comes down to three big steps:

1. Define a structure for my notes (build a knowledge model).

2. Create rules for how information flows in and out (protocols for adding new knowledge entries and retrieving old ones).

3. Build connections between notes (so it becomes more than just a library — something I can use for real sensemaking and solving complex tasks).

Following this roadmap will eventually allow me to convert my notes into a machine-readable format (e.g. choosing the right database and retrieval method so AI can start working with them the same way I do).

That’s the first real step toward self-automation.

So let’s begin where it all starts: by analysing my own messy collection of information — the raw material I want to turn into what many people call PKM, or Personal Knowledge Management.

The Current State of My Notes

Right now, my “knowledge base” is nothing more than 1,000+ scattered notes which I’ve collected over past 2 years. No structure. No system.

Which means that when I actually need to use them (whether for work projects, writing, or even planning a trip) I can rarely find what I’m looking for.

That was a sobering moment.

I caught myself asking: “If I can’t retrieve what I need, what’s the point of taking notes at all?”

That’s when I realised: to truly navigate my externalised mind palace, I need a map. A structure that lets me move quickly between rooms, find the right shelf, and retrieve the right piece of knowledge.

And defining that structure? That’s the very first step on this journey.

Step 1: Building a Knowledge Model — the Backbone of My External Brain

If I want to transform my scattered notes into a structured knowledge base, I need more than folders and tags. I need a knowledge engineering framework which is a model that gives shape to chaos.

For this, I turned to the idea of ontologies.

Think of an ontology as a mental framework, a schema that helps you organise and interpret information. (The roots of this go back to Bartlett’s Schema Theory in 1932.)

At its core, a schema is a logical blueprint. It defines the features of knowledge: concepts, hierarchies, properties, and the relationships between them.

Here’s how I’m adapting that blueprint for my own notes:

• Concepts: mind spaces, tasks, quotes, ideas, reflections

• Containers & hierarchies: folders → subfolders → notes

• Labels: tags that act as metadata for analysis and retrieval

This structure will be the backbone of my external brain — a way to classify, connect, and navigate my notes across different topics and projects.

Step 2: Create Rules for How Information Flows In and Out — Ensuring Consistency in PKM

A structure by itself isn’t enough.

If I organise my existing notes but then go back to adding new ones in the same chaotic way I did before, the structure won’t survive.

It’ll collapse under the weight of inconsistency. (That’s exactly what happened with my early PKM experiments I wrote about earlier.)

So what was missing?

A schema.

A schema acts like a map of my mind palace:

1. When I add a new piece of knowledge (a data entry), the schema helps me get into the “right headspace,” find the right container of knowledge, and annotate the information with a standard set of metadata.

2. When I need to find and extract existing knowledge, I use the same schema for navigation.

Think of it as a grid map that helps you take shortcuts to the exact information you need to retrieve.

And here’s the surprising part: the schema has even more to offer.

Step 3: Creating Relationships Between Notes — the Key to Making Them Usable

Beyond serving as a map for navigating my mind palace, the schema also works as a guide for two critical things:

• A playbook for classification: defining the type and purpose of each note, assigning it to the right container (folder), and tagging it with the right metadata.

• A rulebook for linking notes: creating interconnections that make the knowledge base dynamic. For example, a note about the Thousand Brains Theory might live in my “AI” research folder (within the “Areas of Interest” mind space), but also link to “Projects” when I use it to design AI solutions, and “Beliefs” when I reference it in a blog post about AI.

If the schema defines the knowledge spaces (containers) and their locations in my mind palace, then the relationships are the network of roads between those spaces.

Thanks to those “roads”, I don’t have to enter every room one by one when retrieving knowledge — I can move faster using shortcuts provided by the network of connections.

This is also how learning through movement happens: as I move from one room in my mind palace to another, I start to see how the entire palace connects, and how it reflects a model of the world itself.

That’s where sensemaking takes place. And this is what makes a knowledge base more than just storage.

It’s the relationships that allow me to pull the right piece of knowledge, at the right time, to solve the task in front of me.

Formalising My Personal Knowledge Retrieval System

Our brains are incredible, but they weren’t built to hold and manage vast amounts of information.

I feel that limitation every time I tackle a task that requires pulling together multiple pieces of knowledge at once.

That’s why my bet is on externalising memory — building a structure that allows me to:

• Know where to find information — e.g., which folders or subfolders contain the relevant note.

• Design knowledge retrieval pipelines — e.g., for preparing a presentation on AI, I might pull in:

  • Insights I’ve recently read about AI

  • Storytelling techniques for shaping those insights into a compelling narrative

  • Visualisation methods for presenting them effectively

• Move knowledge into working memory — extracting what I need from “long-term memory” (aka my Second Brain) and loading it into my “working memory“ (i.e. “short-term workspace”), where I can start applying it directly to the task.

Before I even think about automating this with AI, I need to make sure my manual knowledge retrieval workflow works reliably.

And that’s exactly what I’ll explore in the next article — putting this system to the test with the real tasks I face every day.

To navigate the physical world, we need a map.

To navigate the world of meanings, we need mind maps.

Mind maps are the stories our brain tells us about the environment around us.

These mind maps guide us toward our goals, just like a geography map helps you travel from point A to point B.

What Is Our Memory?

Maps can be catalogued into an atlas.

Your atlas of mind maps is essentially a representation of your knowledge about the world.

Your memory is where these mind maps live. It’s your mental “mind palace.”

How We Use Memory to Solve Tasks

Achieving any goal (personal or professional) requires solving tasks.

Solving tasks requires thinking.

Thinking, in turn, relies on:

• Retrieving knowledge (mind maps) relevant to the task — stored in long-term memory

• Processing incoming inputs from your senses — handled by short-term memory

Why Long-Term Memory Matters

The better your maps, the better you can solve problems.

Your atlas (the long-term memory) is reusable, evergreen knowledge your brain taps into every time you face a challenge.

According to Tulving’s theory, we use three types of long-term memory when solving tasks:

1. Semantic memory: Knowledge about the subject matter domain (definitions, concepts, rules — e.g., the legal domain).

2. Episodic memory: Records of past experiences with similar tasks (what worked, what didn’t) so we don’t repeat mistakes.

3. Procedural memory: Step-by-step workflows for completing a specific type of task.

Why AI Agents Need Memory Too

AI solves tasks in a surprisingly similar way to the human brain:

• We give it a prompt — the situational input, like our short-term memory.

• AI retrieves facts and knowledge from its pre-trained data (its “long-term memory”).

• It combines both to generate an output.

The inefficiency?

Pre-trained data is a black box. Often it’s too generic to deliver personalised, high-quality outputs.

But what if AI could tap into your memory (your own curated, structured knowledge) when solving tasks?

Digitalising My Memory for AI

In my previous article, I shared some of my experiments with PKM (Personal Knowledge Management). But when I started exploring how to connect PKM to AI, I realised something was missing.

It was a map that AI would need when tapping into my externalised mind palace.

A map that will help AI to find the right “room” in my mind palace to retrieve relevant knowledge.

The concept of memory gave me that map.

So I rebuilt my PKM system using:

• the distinction between short-term and long-term memory

• Tulving’s theory for organising semantic, episodic, and procedural knowledge

In the next article, I’ll share the results of this experiment and explore how digitalising your memory can transform AI from a generic assistant into a true extension of your mind.

Stay tuned.

AI has extraordinary creative power.

But to fully harness it, you need more than good tools - you need good prompts. And good prompts start with context.

High-quality prompting isn’t just about syntax. It’s about knowing what you want to achieve (your vision of success) and how you want it to look (your taste). In other words: you need to engineer the context for the AI to work with.

But here’s the problem: most of this context lives only inside your head. And unless you’ve structured it in a way AI can access and understand, you’ll keep getting generic, low-value outputs.

That’s what this article is about: understanding how to externalize and structure context so that AI can generate results you actually want.

Sufficient Context = High-Quality Output

You’ve heard the rule: garbage in, garbage out.

It applies more than ever in the age of generative AI. Without context, your prompts are just noise. And the output will be, too.

To give AI sufficient context, you first need to understand your own goals and intentions. And that’s not as easy as it sounds.

Try it yourself: the next time you’re drafting an email or building a slide deck with AI, ask yourself:

• What am I trying to achieve?

• What does “good” look like to me?

• What information would I need to make this decision?

You’ll quickly realize how hard it is to translate that internal clarity into an external prompt. That’s the essence of context engineering.

Context Engineering Takes Time

I’ve been through this too.

Before submitting a prompt, I’d spend 15 to 20 minutes just trying to figure out what I actually wanted and how to express it.

Eventually, I realized I couldn’t keep doing this from scratch every time. I needed to design myself first: to make my goals, preferences, and thought patterns visible and accessible not just to me, but to my AI tools.

Otherwise, every prompt would start with an existential crisis.

Searching for Ways to Reduce Context Engineering Time

In early 2023, I began building my first “context library.”

I started collecting references: snippets from books, social media posts, client feedback. And dumping them into Apple Notes.

By summer 2024, I had over 1,000 notes. But I was barely using them.

That’s when it hit me: collecting is not the same as retrieving. I needed a system to help me find, update, and apply these fragments of context when I needed them.

That’s when I turned to Personal Knowledge Management (PKM).

How PKM Saves Time When Prompting with AI

If context engineering is the new bottleneck, PKM is the new unlock.

I tested several PKM systems:

• ACE by Nick Milo (a structure for mapping ideas)

• Zettelkasten (a classic method for organizing connected thoughts)

• PARA by Tiago Forte (Projects, Areas, Resources, Archives)

My early system included “evergreen” index notes (basically personal knowledge hubs) and content notes with actual highlights, references, or insights.

By the summer of 2024, I had a working PKM setup inside Apple Notes. I had saved over 500 new notes, mostly from reading, conversations, and project debriefs.

A year later, something changed:

• ~30% of my notes had been reused and updated.

• I was using them regularly in prompts and workflows.

• My speed and output quality improved significantly.

But even then, something was missing.

Designing a PKM System for AI Context Retrieval

Most PKM systems were built in the pre-AI era. They’re optimized for human reading and writing - but not for machine assistance.

They don’t include any structure for your identity: your taste, tone, decision-making patterns, or strategic goals. Why would they? In the past, you didn’t need to teach a machine who you were.

But now you do.

That’s why I’m building a new kind of PKM: a system for designing a digital version of yourself.

This isn’t just note-taking. It’s identity modelling.

I want AI to understand me the way a creative partner would.

To get there, I’m building a system that captures:

• My preferences

• My knowledge map

• My style and voice

• My strategic goals

In the next article, I’ll show you how I’m doing it.

Stay tuned.

There’s a lot of noise out there when it comes to AI. Everyone’s trying to use it for everything.

I’ve been there too. As a legal professional, I tested every new AI tool that landed in my feed.

At one point, my day job basically became a series of chaotic experiments: reading tweets, jumping into new tools, and attempting to retrofit them into my workflows.

The result? I was overwhelmed. My productivity tanked. I was chasing novelty, not impact.

That’s when I realized: if I want to make AI work for me, I need a system, a framework for deciding if an AI use case is worth my time, energy, and budget.

This article shares the foundation of that system.

Cutting Through the Noise: A Use Case Filter for Knowledge Workers

The question I started asking wasn’t "what can AI do?”.

It was: what’s worth automating based on real ROI?

And more importantly: how do I know if the AI use case in front of me is the right one for the type of task I’m trying to automate?

This is the mindset shift that led me to build a framework.

It helps me:

• Filter the market noise

• Avoid wasting hours chasing novelty

• Choose automation use cases that lead to measurable returns

Let’s walk through the key concepts.

The 3 ROI Drivers Behind AI Automation

When evaluating an AI use case, I look at its potential to deliver ROI in one or more of three ways:

• Speed: Can it reduce turnaround time?

• Cost-efficiency: Can it lower my resource use or increase my margin?

• High Quality: Can it solve complex problems without errors?

Important note: you rarely get all three.

That magical use case that delivers fast, cheap, AND flawless work?

It doesn’t exist (yet). In 99% of cases, you’re choosing two out of three.

So the real question becomes: which two matter most for the task you’re automating?

Mapping the Right AI Use Case to the Right Type of Task

To answer this, I started mapping knowledge work based on task complexity (a scoring system I introduced in the last article) and the AI implementation model used to automate it.

I borrowed the structure of the Ansoff Matrix to compare four knowledge work models:

1. Without AI:

• Freelance expert (simple tasks)

• In-house firm expert (complex tasks)

2. With AI:

• AI Agent + Expert-in-the-loop (simple tasks)

• Expert + AI Assistant (complex tasks)

Old Way: Solving Tasks Without AI

Before diving into the new AI-powered approaches, let’s look at how tasks were traditionally handled:

1. Freelance Experts from Marketplaces

• Best for low-complexity, repeatable tasks

• ROI focus: Cost-efficiency (but not speed or quality)

2. In-House Experts from Firms

• Best for high-complexity, high-risk tasks

• ROI focus: Quality (but not cost or speed)

The traditional options only deliver on one ROI factor.

New Way: AI-Powered Knowledge Work Models

Now let’s explore how these two types of tasks are handled when powered by AI:

1. AI Agent + Expert-in-the-Loop

• Ideal for low-complexity tasks (e.g. compliance checks, LinkedIn posts)

• ROI: Speed + Cost-efficiency

2. Expert + AI Assistant

• Ideal for high-complexity tasks (e.g. legal memo, client proposals)

• ROI: Speed + High Quality

This is where AI makes a real difference: it helps unlock two ROI dimensions instead of one.

Conclusion: Choose Your Use Case Based on Task Type

In this article, we identified the two key implementation models for AI:

1. AI Agent (automation-first, expert-in-the-loop)

2. AI Assistant (expert-led, automation-supported)

Knowing which model fits your task type - and what ROI to expect - is essential to avoid wasted time, effort, and hype fatigue.

Use this framework to decide which AI use case best fits your task based on frequency, complexity, and ROI potential.

In the next post, I’ll walk through real case studies from my own work, applying both use cases to automate actual tasks.

Stay tuned.

In the previous article, I shared a playbook for task mapping - a method to gain a “big picture” of how tasks function within a micro-firm operating in the professional services sector.

Now, it's time to move from mapping to scoring.

The next step in the automation journey is task assessment: identifying which mapped tasks are not only automatable but also worth automating.

That means answering on two questions:

• How hard is it to automate the task and how much would it cost?

• What’s the payoff if I do?

The Goal of Task Assessment

The goal of task assessment is to estimate:

1. The resources required to automate a task (based on its complexity)

2. The expected return from investing in that automation

This scoring system helps prioritize tasks in an automation pipeline.

But first, I need a consistent method to evaluate them.

Understanding the Anatomy of a Task

Every task has two key components:

• the Outcome it’s meant to deliver

• the Workflow required to produce that outcome

Since automation targets workflows, that’s what we’ll assess.

Stage 1: Task Complexity Assessment

To assess complexity, I break down workflows using four dimensions:

1. Knowledge Domains Involved: Does the task rely on one domain or multiple?

2. Workflow Structure: Is it a simple checklist or a multi-scenario decision tree?

3. Data Sources Required: Few internal docs, or many external + internal sources?

4. Output Modality: Is the output unimodal (e.g., just text) or multimodal (e.g., text + visuals + video)?

I use an X-shaped matrix to plot these variables and assign each workflow a complexity score (ranging from 2 to 8).

Workflows with higher scores are more complex - and more resource-intensive to automate.

Example: Complexity Assessment of Two Marketing Tasks

Let’s assess two real-world tasks from the "Marketing" function:

• Task 1: Write a short LinkedIn post

• Task 2: Create an SEO-optimized blog article with visuals

Step 1: Map out the workflow for each task.

Step 2: Score each workflow based on the 4 complexity dimensions.

Step 3: Use the scoring matrix to determine the complexity category (Simple / Complex).

This gives me a clear picture of how much time, money, and expertise each task might require to automate.

Stage 2: Estimating Automation Resources in $

Once I’ve scored a task’s complexity, I can estimate:

• The domain experts needed

• The required tech stack (e.g., GPT, Zapier, Notion, n8n)

• Data integrations and API costs

• Output formatting tools

For example, for a LinkedIn post automation I need:

• one expert domain (SMM)

• 5-step linear workflow

• two static content sources (company’s messaging guide, LI guidelines)

• text-only output

This likely requires:

• a simple no-code automation setup (Zapier or n8n)

• input of one expert to set up prompt structure and enhance it as we go

• subscription/API access to LLM and posting tools

Each of these inputs can be priced out to give a dollar-based investment estimate.

The more complex the task, the more investment is needed.

Why This Matters Before Calculating ROI

You can’t calculate return on investment unless you first understand the investment.

This playbook gives me a structure to:

• score task complexity

• estimate the time and budget required for automation

In the next article, I’ll cover the final step: assessing impact and return - so I can make decisions which tasks to prioritise for automation based on actual ROI.

Conclusion: From Score to Strategy

Before you automate, assess.

Task assessment allows me to identify where to focus my efforts - and where automation will pay off the most.

I now have the tools to score complexity, estimate costs, and prepare for ROI analysis.

Next up: Task Automation Impact Assessment - the final post in this 3-part Task Audit series.

Stay tuned!

It’s been four months since I started my Substack, documenting my self-automation journey.

In the last article, I outlined the principles that guide my automation approach. But as I started applying those principles, one question kept surfacing:

What tasks should I start my automation with?

Before building an AI workforce to replace myself in business, I needed to clearly define the core tasks that I (and other knowledge workers) handle daily - tasks that either consume my own time (creating a speed bottleneck) or require outsourcing and delegation (creating a budget bottleneck).

The goal is not merely to reduce time or costs, but to strategically redirect freed-up resources toward business growth - until I achieve what Buckminster Fuller called "ephemeralization".

This strategy is crucial for measuring a clear ROI on automation.

Process-Centric vs. Outcome-Centric Tasks

Not every task is worth automating.

Some tasks are fulfilling and creative - like choosing a gift for a friend. Let’s call it process-centric tasks - the kinds of tasks where the experience and process matter as much as the outcome.

But then there are outcome-centric tasks - the ones where speed and efficiency matter more than the process itself.

In business, these tasks dominate (with some exceptions like sales calls, face-to-face consulting etc).

That’s where I envisioned the ultimate objective of my self-automation project:

But first, I needed to define the profile of the knowledge worker whose tasks I aimed to automate and map out their workflows.

Defining the Persona: The Solopreneur Knowledge Worker

Before identifying tasks for automation, I needed to clarify the persona whose tasks I would focus on.

After reflecting on my own journey and consulting experience (read “The Existential Crisis of Knowledge Worker” to find more), the following profile emerged:

• service-centric Knowledge Workers (professionals delivering specialized expertise in sectors like legal, consulting, accounting, engineering, research)...

• …who operate as Solopreneurs (those operating as one-person firms or micro-agencies with fewer than 10 employees).

Why service-centric roles?

These roles are knowledge-heavy, requiring specialised information and expert reasoning to provide value to the customers.

That’s where AI has one of the strongest use cases.

Why solopreneurs and micro firms?

They’re the ones who stand to gain the most from AI automation, as AI-driven workflows can transform a one-person consultancy into a high-margin, scalable operation without additional hires.

It’s also a substantial market ripe for self-automation solutions.

There are 750k micro-consultancy firms in the UK alone, making up 14% of all businesses.

In the US, this number jumps to 2.5 million.

Mapping the Knowledge Worker’s Tasks

Now that I’ve defined the target persona, it’s time to map out their key tasks.

Knowledge-centric businesses operate like systems, with four core functions:

1. Marketing – Generating inbound leads

2. BD & Sales – Generating outbound leads and converting all leads into clients through qualification calls, proposals and negotiation

3. Delivery – Providing expert services and deliverables

4. Admin – Managing finance, legal, IT and backoffice processes

Each function contains role(s), each with its own set of goals and tasks.

Each task should result in an outcome that directly contributes to achieving the overall goal of the business (e.g. revenue generation).

Here is how it looks for the marketing function:

• Role: Marketing Manager

• Goal: Inbound Leads

• Task: Blogging on LinkedIn

• Outcomes: Published blog posts.

This categorization and hierarchy will be crucial for ontology modelling and knowledge extraction as we automate knowledge-intensive tasks.

Creating Task Automation Pipeline

Now that the task map is in place, the next step is to create task automation pipeline - the specific set of tasks that:

• Have clear, repeatable outcomes (i.e. outcome-centric tasks)

• Consume the most of my personal time or require delegation or outsourcing

• Create bottlenecks for revenue growth

In the next article, I’ll share how I applied this audit to my own work, identifying the specific task I chose to automate first and why.

Conclusion: Building the AI-Powered Business Operating System (AI BOS)

In this article we’ve mapped out the core functions, roles, and tasks of the solopreneur knowledge worker.

But identifying tasks is just the first step.

The next step is to connect these tasks into a unified orchestration layer - a network of AI-powered workflows that work together to form a cohesive AI-powered Business Operating System (AI BOS).

This AI BOS will not only automate individual tasks but also align them strategically to drive sustainable business growth autonomously.

Stay tuned!

Automation is a time-intensive endeavor.

And in the realm of self-automation, my most valuable resources are time and focus, the very assets that can make or break my efforts.

Before diving into automation, I had to define the principles that would guide my approach.

Why? Because automation is only valuable if it achieves a clear, measurable outcome. Otherwise, it’s just busywork disguised as productivity.

So, here are the principles that will shape my self-automation journey. They’re designed not just to streamline my work but to ensure the workflows I create are scalable, explainable, and, most importantly, ROI-driven.

Principle 1: Objective-Focused Automation

I don’t automate tasks for the sake of automation.

I start by clearly defining a success metric and ROI for every task. This is the only way to know whether automation is truly effective.

For example, before automating the process of LinkedIn post creation, I asked myself: “What outcome am I trying to achieve here?” The answer was simple: I want to reduce the idea-to-post time by 50% - a clear, measurable outcome.

How do I determine which tasks to automate first?

I audit existing workflows to identify the highest-impact, most repetitive tasks. This approach keeps my efforts laser-focused on what truly moves the needle.

Principle 2: “Know-how“ Knowledge Retrieval

Not all data is useful.

In fact, raw data is meaningless until it’s been processed and applied in context. That’s why I prioritize extracting practical knowledge - the kind that directly contributes to achieving the defined outcome.

For me, practical knowledge is more than information.

It’s data that has been:

• Captured systematically

• Applied to real-world tasks to build "know-how"

• Structured for easy retrieval and reuse

In other words, if the data doesn’t directly connect to a tangible task or workflow, it’s just noise. By focusing on practical knowledge, I keep my systems lean and purposeful.

Principle 3: Explainability and Verifiability

No black boxes allowed.

If I can’t explain how a workflow functions or if my team can’t verify its logic - it’s not scalable.

Every workflow I build must be:

• Explainable: Clearly documented, with structured knowledge that anyone can understand

• Verifiable: Designed with logic that can be audited by third parties

Why is this critical?

Because at some point, I’ll need to delegate the maintenance of these workflows to others.

If the workflow isn’t transparent, it’s a liability—not an asset.

Principle 4: No-Code Implementation

Today automation doesn’t require advanced coding skills.

For this self-automation journey, I’ll be using no-code tools to create AI-powered workflows. This not only keeps the process accessible but also ensures that the workflows can be iterated rapidly without technical bottlenecks.

What does this mean for other solopreneurs and knowledge workers?

If you’re not a developer, you can still implement AI automation without writing a single line of code.

And in the upcoming posts, I’ll show you exactly how.

Connecting the Dots: Why These Principles?

• The first two principles prevent me from wasting time on tasks that don’t have clear success metrics or tangible ROI.

• The last two principles ensure that my automation workflows are scalable and transferable—from a knowledge perspective (explainability) and a technical perspective (no-code tools).

Bonus Insight: Automation as a Cognitive Exercise

There’s an unexpected side effect of diving deep into workflow automation.

It forces my brain to engage in complex problem-solving.

According to a Microsoft study, passive reliance on AI can actually reduce cognitive effort and confidence.

But actively designing and implementing AI workflows?

That’s like a gym workout for your mind.

Conclusion

With the roadmap to self-automation laid out and guiding principles clearly defined, it’s time to get practical.

In the next post, I’ll dive into Task Auditing—the step-by-step process of identifying the most impactful tasks to automate.

Stay tuned.

First-principles thinking teaches us that complex projects should be broken down into smaller, foundational tasks.

My self-automation challenge is one such complex project, with multiple possible approaches and outcomes.

To accomplish it effectively, I need two key elements:

1. a roadmap that breaks down the tasks and sequences them chronologically, and

2. foundational principles to guide the direction.

This article will explore the roadmap, while I'll cover the guiding principles in the next one.

Roadmapping Exercise

The goal of my roadmapping exercise is straightforward: define a starting point and create a work plan to achieve desired results.

For my self-automation roadmap, I've chosen to work backwards:

1. First, I identify 2-3 most frequent tasks from my weekly task list that I want to automate

2. Second, I envision my ideal outcome for each task

3. Third, I visualize a workflow that will generate the desired outcome autonomously and repeatedly

4. Finally, I reverse engineer the steps to build the workflow and identify the required resources for its operation

This backward-planning approach helps me navigate all my complex projects, including my self-automation project.

Below we will explore each of these steps in detail.

Step 1: Defining Repeatable Tasks for AI Automation

When planning my self-automation project, I begin by envisioning a clear, objective-driven outcome.

My primary goal is to automate my daily routine, so the first step is identifying the specific tasks that make up a significant portion of my workday.

I've chosen to focus on tasks that meet these criteria:

1. tasks with high frequency (occurring at least weekly)

2. tasks that produce text-based deliverables (documents, emails, messenger responses)

3. tasks that require retrieving knowledge from my experience and personal involvement

Here are some examples of these tasks:

1. creating content for personal blog and social media

2. building tailored proposals for enterprise clients

3. providing feedback for candidates from our hiring pipeline

While the list of tasks I want to automate is longer, these three tasks alone consume about 20% of my weekly work time.

Step 2: Imagining ideal outcomes of AI Automation

My strategic goal for the next 6–12 months is to develop and implement task automation that will save at least 20% of my weekly working time—effectively enabling a four-day work week for me.

Here are examples of ideal outcomes I expect my personal AI Assistants to generate autonomously:

1. A ready-to-publish LinkedIn post—instead of sharing ideas with our content team (input) and reviewing drafts, I can dictate to my AI assistant and receive a finished post (output)

2. A customized enterprise client proposal—after a sales call, I can upload the transcript (input) to my AI assistant, which will generate a tailored proposal (output) by retrieving relevant information from the database of services according to the client's needs identified during the call

3. Candidate feedback for our hiring pipeline—rather than waiting days for feedback, our recruitment team can send candidate information to my personal AI assistant (input) and receive responses as if from "real Nestor" (output)

I'll share many more use cases for personal AI assistants in upcoming articles.

Step 3: Visualizing the AI Automation Workflow

After identifying tasks to automate and imagining ideal AI-generated outcomes, I focus on visualizing how I would interact with my self-automation system to receive these AI-powered results.

Here is the workflow I envisioned:

1. Launch a specialized AI Assistant focused on a specific task—an application that understands my expected outcomes and operates within a defined domain of my knowledge base

2. Input data (preferably by voice) and select an appropriate prompt from pre-designed prompt catalogs (each specialized AI Assistant has its own catalog designed for specific tasks within its domain)

3. Wait while the AI Assistant's built-in LLM generates an initial output, retrieves relevant data from my personal knowledge base to enhance it, and delivers it to me in a human-readable format

From a user's perspective, the AI Assistant will function through a chat interface where I can interact with my externalized knowledge and complete tasks in seconds.

My ideal scenario includes sharing my personal AI assistants with colleagues who need my input, allowing such AI versions of myself (aka "Nestor AI") to respond instead of "real Nestor."

Step 4: Structuring Work Plan for Self-Automation

Now that I've defined my automation tasks, outlined outcomes, and designed a high-level workflow, I can create a roadmap and determine where to begin this complex project.

My roadmap to self-automation consists of two core stages:

Stage 1: Building a Knowledge Layer as a Foundation for Self-Automation:

1. Set up a personal note-taking system to capture information (tasks, logs, projects, areas)

2. Implement techniques to convert information into knowledge (taxonomy, ontology, playbooks, checklists, and tracking systems)

3. Make personal knowledge machine-readable by choosing a framework to transform unstructured information into a structured database

Stage 2: Implementing Automation:

1. Select initial repetitive tasks that require knowledge retrieval from my personal knowledge base and allocate specific knowledge for each task

2. Transform target tasks into prompts and identify the most effective LLMs for each type

3. Deploy LLMs, connect them to my machine-readable knowledge base, and implement automated task solutions through knowledge retrieval and agent actions

Why Not Use Existing AI Solutions?

You might ask, "Nestor, why are you doing this? Why not use one of the countless existing AI apps and agents for automating personal tasks?"

Why?

Because they're all essentially black boxes:

• I can't verify which knowledge bases these AI solution providers used to build them

• I don't know which LLMs they're using to generate outputs

• These unknowns force me to spend excessive time on prompt engineering to get quality results

• I can't fully trust them with my knowledge (which is my proprietary IP) or my inputs (which often contain personal and sensitive data)

Moreover, existing LLMs have limited memory for long-term use. Once you hit this limit, you must rebuild the AI app's context from scratch—which explains why many AI apps experience strong initial growth followed by steep decline.

Summary

In this article, I've outlined my roadmap to self-automation—a plan that will likely need adjustments as we progress alongside industry developments.

Yet having a detailed roadmap alone isn't sufficient for such a complex project.

The sheer number of approaches, concepts, and frameworks for self-automation creates a challenging landscape to navigate.

This is precisely why I need foundational principles to act as my compass, guiding my progress and focus throughout each stage of knowledge structuring and AI automation.

In the next article, I'll share these principles, and then we'll be fully equipped for our journey into self-automation.

Stay tuned!

Every automation needs to be objective-focused.

To create objective-focused automation, we must clearly visualize our desired result. This visualization needs to be detailed and specific—otherwise, the automation won't succeed.

I aim to keep my personal blog as practical as possible.

For every automation experiment I write about on my Substack, I’ll begin by describing the specific objective I'm trying to achieve to avoid generic and abstract discussions.

In this blog post, I'll show how automation drove our sales growth.

My Starting Objective

Last November, I set a clear goal for Legal Nodes: achieving 40% month-over-month growth in sales. I created a detailed vision with specific numbers.

At the time, our sales growth was inconsistent, averaging 10-20% MoM. I decided to target 40% for several reasons:

1. 40% was a solid benchmark for Legal Nodes' stage of development (YC recommends this target too)

2. We started preparing for a new funding round, and growth rate became a key metric

3. I drew inspiration from Deel (2022's fastest-growing company) which specializes in cross-border employment—parallel to our focus on cross-border corporate structuring & compliance

With the goal set, I faced crucial questions: What's next? Where do I start? How do I identify the obstacles between the current sales metrics and 40% MoM growth?

I realized this challenge needed a fresh thinking approach, since previous growth initiatives hadn't succeeded.

Every Problem Starts With It

Every organization is a system.

If there's a problem in the system and you're part of that system—you're part of the problem.

When I first read this phrase in the book "Liminal Thinking," I realized that I was likely part of our unsatisfactory sales growth problem since I was part of the sales system.

That's where I found the answer to my question "Where to start"—I had to start with myself, or more precisely, with how I thought about our sales system.

But immediately I faced a new challenge: How do I examine my own thinking?

A Challenge of System Thinking

Every system consists of variables and the relationships between them. The more variables there are, the more complex the system becomes.

When I started analyzing my understanding of our sales system, I realized that the number of variables (ICP & JTBD, PMF, sales funnel, etc) and their interdependencies were so numerous that my brain couldn't process them all at once.

I found myself stuck, wondering "How can I get all these elements of the sales system out of my head to see the complete picture?"

The solution I found that worked perfectly—a whiteboard.

The Tool That Makes Complex Information Digestible

As David Allen said in his GTD book: "Our brain isn't designed for holding information. It's designed for processing it and generating ideas."

So when dealing with complex systems that have multiple variables and relationships, it's essential to see the whole picture before generating ideas and making decisions.

This requires getting the information out of your head and onto paper first.

Whiteboarding helped me organize complex information in a way my brain could digest.

(It was also a pleasant surprise when I realized that visualized information wasn't just helpful for my own thinking, but also for communicating with team members and creating shared visions of what we want to achieve at Legal Nodes).

This Skill Unlocked the Next Level of Thinking

When I started whiteboarding last November, my diagrams were unreadable and my attempts at visualizing systems were incomprehensible to team members.

I had no prior drawing experience or visualization skills.

I spent two months researching this topic, practicing visual communication with the team, and buying every book I could find about information design.

Earlier this month, my wife began complaining about the lack of space on our bookshelf because of the number of information design books I'd accumulated over the past three months.

But it was worth it.

The basic visualization skills I developed helped me to:

• Organize system information externally (outside my head) through charts, diagrams, and decision trees;

• Identify discrepancies and stuck patterns in the system;

• Quickly iterate on the system and share changes with the team through visual communication.

But there's more.

Whiteboarding truly proved its value when I combined it with the ITTTF thinking model—a framework that helped make the system autonomous.

The Final Puzzle Piece for Systems Thinking (and unlocking growth)

Information alone isn't knowledge.

But when information is processed into a "How to" format, our brain transforms it into knowledge.

The ITTTF (if-this-then-that-framework) became my daily tool for converting information into knowledge. Combined with whiteboarding and visualization skills, I finally understood how to analyze, communicate, and scale systems.

And that's when our sales growth took off.

The visualized ITTTF framework, now central to our sales system, had a profound impact on our team. It unified our approach and helped all team members align around a shared vision for our 40% growth target, channeling our collective energy toward achieving it.

But Where is Automation in All of This?

Implementing automation requires a system to reach a certain level of maturity before AI can effectively reduce manual work.

I realized this after several unsuccessful attempts at sales process automation.

It became clear that I first needed to understand how the system should look before starting AI automation experiments. In other words, I needed to determine what maturity criteria a system should meet to be ready for AI.

I've identified these key criteria for system maturity:

• The system operates autonomously—team members run and maintain it independently since they share both a clear vision of the goals and a common understanding of the system's structure.

• The system evolves dynamically—team members understand its original design, feedback loops are defined, and iteration cycles are established—which keeps every update of the visualized system clear and accessible.

• The system is self-regulated—team members can adjust it without the founder's input because they understand both the current structure and the principles behind its design.

The key lesson I learned is that designing a dynamic and autonomous system requires proper thinking tools and the skills to externalize such thinking—structuring it outside the brain so it can be effectively communicated to the team.

What’s Next?

This was my key insight while updating our sales system:

automation must first happen in our minds.

To achieve this, you need to:

• externalize information from your mind

• organize it into variables, dependencies, and desired outcomes

• create a visual system that answers "How to" questions (converting information into knowledge)

• share this with your team and establish feedback loops for continuous improvement

While this isn't automation in the conventional sense, it makes the system scalable by streamlining the thinking process of team members who operate it.

It also unlocks the possibility to proceed to the more exciting part: AI automation experiments that we can run using this system foundation.

I'll share more about this in upcoming blog posts.

Stay tuned!

Midway through my 10-years legal and tech career (which I touched on in my previous post), I realized what it truly means to be a "knowledge worker." I observed how accumulating industry and legal knowledge was advancing my career as a legal practitioner and tech entrepreneur.

However, during a meeting with one YC founder in 2020, I heard a provocative question: "Nestor, it's great to see how your legal and industry knowledge is evolving, but how do you plan to scale it?"

This question struck me deeply at the time, but the pandemic and the full-scale war in Ukraine in the following years shifted my focus to mere survival.

The question from my fellow founder resurfaced in late 2023 when I received a recap from Fireflies AI analyzing over 1,500 meetings and nearly 100,000 minutes of work conversations that year. The statistics were shocking.

After seeing these shocking statistics, I asked myself an existential question for the first time: did I want to spend my next 20-25 years of active professional life this way?

Or was there another path?

It began with innocent note-taking..

Ten years ago, as a law school student, I began personal note-taking. I collected interesting idioms from fiction books to impress friends in conversations (an embarrassing habit, I know 🤦‍♂️)

A couple of years later, I expanded my note-taking to personal projects. One of my first projects was building a family tree—I interviewed over 20 relatives to map our family connections and even reconnected with distant relatives whose ties were lost during WWII.

Back to that time my personal database of notes was growing rapidly, and I felt proud of my work—until I realized there was no correlation between the size of my note collection and my personal productivity.

That was the moment I first faced the challenge of processing unstructured information.

Painful Lesson Learned

In 2022, my personal note collection had grown so large that I struggled to find anything in it. I began forgetting what information I had already added, leading to duplicate notes. Projects and tasks I tried to track became a complete mess. And my ideas folder? It transformed into a graveyard of unrealized dreams and thoughts.

During this time, I started losing interest in personal note-taking until I encountered this phrase: "Information is not knowledge"

It struck me like lightning.

That's when I realized why intense note-taking alone wasn't making me smarter or more productive—a painful insight, given the mountain of notes I had accumulated.

The phrase also triggered a memory from my law school philosophy classes about epistemology.

I decided to revisit those university lectures about knowledge and research current solutions in this space. That's how I discovered the world of knowledge management systems (aka "second brain").

When My Note-Taking Evolved Into a Second Brain

2023 became a year when I started focusing my note-taking not just on techniques of capturing information but also on a methods and thinking tools which help to transform captured information into actionable knowledge (reading epistemological papers and books about general semantics, exploring knowledge graphs).

Over time, I added more categories of information to my personal notes:

• ideas

• meeting notes

• self-tracking projects

• personal CRM

• professional data

• checklists

• templates

• travel guides

• and much more

As I expanded the types of information, I experimented with various note-taking systems (like Zettelkasten, PARA, and ACE) and different digital notebooks (Apple Notes, OneNote, Notion, Obsidian, and even Roam).

These initial note-taking experiments ultimately led me to the creation of a personal knowledge graph I started building last year.

That shift changed my note-taking for the better. Finally, my decade of collecting wisdom from books, movies, podcasts, and conversations with smart people began working for me, enhancing my decision-making, productivity, and creativity.

Then came another exciting revelation.

The spark of an idea to tackle knowledge work challenge

When I decided to commit to creating a PKM system, AI and LLMs weren't yet in the spotlight. At the time, I simply thought that having a formalized personal knowledge base would help improve my productivity in professional and personal life.

Then AI boomed in early 2023, striking me with the answer to my YC friend's 2020 question about scaling myself.

I realized that AI was the final piece in solving the puzzle of how knowledge workers like myself could scale.

This hypothesis I developed in 2023 was later validated by the market narrative of verticalized AI agents, which gained strong support from major VCs and big tech companies in 2024.

That was when my curiosity sparked a new question: "What if I connect AI to my second brain?" This led to another thought: "How should I structure my personal knowledge to make it readable by both humans and AI?"

AI can’t automate your life if it’s not connected to your second brain

Today there are many discussions about personal AI agents that will handle our daily routines in the nearest future:

• you won't need a personal psychotherapist—AI agents will help you self-reflect and identify mental barriers blocking your personal growth;

• you won't need a personal gym coach or nutritionist—AI agents will create training programs and meal plans while tracking your performance;

• you won't need a personal accountant—AI agents will prepare and submit your annual financial statements.

I support these predictions, but with one condition:

These exciting AI possibilities can only become reality for people who are disciplined enough to formalize their personal knowledge about their mental and physical health, habits, and belief systems—so the AI tools can be connected to a person's "second brain" and use this information for producing outcomes.

Without this foundation, we'll only see unhelpful AI applications that provide overly general, impersonal advice.

So what's next in this story?

If you're still here (and I know this has been quite a long read), thank you for staying with me through this personal reflection!

In this blog post, I want to share my next big professional and personal milestone—as a reminder for my future self.

It focuses on experimenting with AI in a professional and personal life—exploring how we can structure personal knowledge "outside the brain" in human- and machine-readable formats.

Throughout this journey, I'll share my learnings and experiments at the intersection of thinking tools, "second brain" systems, and AI.

I'm always happy to get feedback and discuss these ideas further. Let me know if you have any thoughts!

This year marks 10 years since I graduated from law school, 7 years since I passed my first bar exam, and 5 years since I founded a venture-backed company.

This milestone has prompted me to reflect deeply and consider "What's next?"

Before planning for the future, it's essential to understand what has been achieved so far. In this blog post, I'll share my reflections on the past decade and outline where I'm headed next.

This turned into a 2-parts long read, but if you're not in the mood for memoirs, there are plenty of photos to browse through :)

My Unconventional Path Into Legal Career

It's fascinating how our first experiences shape our perception and paradigm for life.

In 2015, after graduating from law school, I got my first job at a legaltech startup called Bitlex. It was one of Ukraine's first online legal solutions.

This first work experience felt like jumping off a cliff, as my daily work went far beyond traditional legal tasks (like consultations and document preparation). As the project's team was small, I was also involved in platform technology (such as support tickets), user experience (since all services were provided through the online platform), and monetization matters. At the time, Bitlex even allowed clients to pay for legal services with Bitcoin, which was worth $300 in 2015.

That was the moment I fell in love with tech and developed my fundamental belief—which I've held for the past decade—that technology can improve the legal industry.

Serial Law & Tech Entrepreneurship

Since then, during the next 10 years, I've worked on three other projects in the sphere of legal innovations. In 2016, my partner and I launched the first legaltech lab in Ukraine, where we've been working on mobile apps and chatbots for different legal use cases (emergency legal help for travelers, document generation for divorce processes, legal entity registration, etc.)

Lexnet was where Patentbot, a chatbot for brand protection, was born in 2017. With Patentbot, we secured our first venture capital funding and participated in a Hong Kong acceleration program. We had ambitious plans to scale into the Chinese market—the world's largest for trademark registration services. However, my journey with Patentbot ended in 2018 when a new venture captured my attention.

At the end of 2018, Margo, Max, and I launched Legal Nodes. Our vision was to create a "decentralized global law firm" where a tech-enabled platform serves as a single hub for global companies, helping them solve multinational legal needs at a fraction of the cost.

We planned to achieve this via the creation of a network of local legal providers with an underlying "collaboration protocol," which helps them in handling these cross-border legal requests.

It's been 5 years since our launch, and it has become a true entrepreneurship journey for me. During this time at Legal Nodes, we won startup competitions and got accepted into the London Techstars acceleration program, raised funding from business angels associated with well-known unicorns like Pipedrive, Wise, and NEAR. Additionally, last year we were accepted into the GoogleAI program.

But there's more to the story than just legaltech ventures. While experimenting with legal innovations, together with legal enthusiasts from VCs, universities, and law firms, we discovered Legal Hackers in early 2017—a global movement of tech-passionate legal professionals headquartered in New York.

Within a couple of months we launched the Kyiv Legal Hackers chapter, which grew to become Eastern Europe's largest legaltech community. Before the pandemic, we built a network of over 1,000 members and organized more than 10 international events, including 4 legaltech conferences and 3 hackathons.

However, you don't think that all the above-mentioned legal initiatives led me to give up my legal profession, do you?

The Existential Question Every Legal Practitioner Faces

While working on various law and tech experiments, I've always believed that creating workable legal innovation requires hands-on experience practicing law.

I also realized that success in the legal profession demands more than just technical expertise in legal drafting and research. A good lawyer must be a trusted advisor who understands their clients' businesses. Given how each industry and market has its own unique characteristics, I had to ask myself: which industry did I want to focus my legal practice on?

The answer emerged naturally.

My Journey into Digital Asset Law

In early 2015, during my final year of law school, I discovered Bitcoin through an article that revealed an interesting fact that in 2014 there was a short period of time when a single company in Ukraine controlled about 45% of global Bitcoin mining. While this concentration posed a potential risk of a 51% attack on the Bitcoin network, it also turned Ukraine into a magnet for talented builders, entrepreneurs, and investors from around the world.

At the time, I saw this as a fascinating niche, recognizing there would likely be many legal questions surrounding Bitcoin's legal status. In 2016, I wrote my first legal article exploring how Bitcoin could be used as authorized capital for companies. I also gave a presentation at one of Kyiv's first Blockchain Conferences and helped organize the first roundtable discussion with crypto industry representatives about Bitcoin regulation in Ukraine.

Ukraine appeared to be an ideal place to begin practicing law in digital assets due to its concentration of talent, which helped me gain deep knowledge about blockchain, consensus algorithms, and encryption technologies.

However, the country's lack of regulatory certainty and rigid civil law system made it challenging to establish legal frameworks for dynamic blockchain innovation. This pushed many crypto legal professionals, including myself, to work with more crypto-friendly jurisdictions abroad.

While remaining based in Kyiv and supporting Ukrainian teams, we frequently traveled internationally to find better regulatory environments for blockchain ventures.

10 Years of Legal Consulting in Digital Assets

Since these early days of my legal practice in the digital asset industry, I passed two bar exams and managed to build extensive legal knowledge in digital asset regulations.

I've consulted over 500 web3 builders, collaborated with crypto-savvy lawyers from more than 20 countries, and discussed web3 legal concepts with GCs and CLOs of leading L1/L2 and DeFi projects. I've also advised regulators in Ukraine, the UAE, and British Overseas Territories, and delivered guest lectures at Geneva University.

So now you know how I got into legal and tech ventures and started practicing law in digital assets.

I'm very happy where this exciting professional journey has taken me, and I'm grateful for all the amazing people I've met and had the chance to work with along the way.

However, midway through my career, I started feeling limited by the knowledge-related aspects of my work. This feeling intensified when the AI boom happened in late 2022, transforming it into an existential challenge.

What is this existential challenge of a knowledge worker?

To be continued in the next post..

Subscribe now