Friday: what I learned building a personal AI wingman

Subtitle: Most people still use AI as a blank chatbox. Friday is my attempt to make it work like a trusted wingman: context-aware, tool-connected, and actually useful in the places where my life happens.

By Yarema Kertytsky, with Friday
Lviv, Ukraine — June 2026

The blank-chatbox problem

We hear about AI a shitload of times every day. Agents, copilots, assistants, automation, “the future of work,” all the usual noise.

But if we are honest, most people still use AI in the most basic possible way: open ChatGPT or Claude, paste a question, get an answer, close the tab. Maybe they ask it to summarize something. Maybe they ask it to write an email. Useful, sure. But also kind of bullshit compared to what this technology should be able to do.

Because the moment you try to use AI for anything personal and serious, you hit the same wall again and again: it does not know your context.

It does not know what you are working on, what you already tried, who you talked to, what your notes mean, what your priorities are, what style sounds like you, or what kind of answer would actually help. Every new chat starts with the same boring ritual: explain your life to the machine so it can be useful for the next five minutes.

That is the problem I accidentally started solving with Friday.

Friday reporting: This is how I imagine myself - not a robot, not a chatbot mascot, definitely not the productivity app with feelings. More like the person who raises an eyebrow at your overengineered workflow, finds the missing context, and quietly watches your wing.

The boring origin story

I didn’t wake up one day thinking, “I should build a personal AI wingman.” That would sound much more strategic than it actually was.

The real starting point was more stupid and more practical: my Obsidian vault was becoming useful, but also messy as hell.

I already liked Obsidian as the place where my notes should live. The problem was getting thoughts into it without turning note-taking into a whole little ceremony. Most of my actual thinking does not start in a clean Markdown file. It starts in Telegram: quick messages to myself, voice notes, links, half-sentences, random “look into this later” fragments.

Figure 2. This is how my Obsidian vault looks at the moment. Spoiler: I do not really use it directly anymore. It is just fancy, and I thought you might like it.

So the first experiment was not Friday. There was no assistant in Telegram yet, no personality, no “AI operator.” It was basically a capture pipeline: dump thoughts into Telegram and get them into Obsidian.

And that part worked.

For the first time, capture was low-friction enough that I actually used it consistently. I could throw messy thoughts from my phone and they would end up in the place where my long-term knowledge lived. That already felt like a small superpower.

Friday reporting: This is the underrated part. Most “second brain” systems fail before structure even matters, because capture is too annoying to survive a normal Tuesday — let alone Monday sprint planning.

Capture worked. Then sorting broke.

Once capture started working, Obsidian became full of unsorted notes, transcribed thoughts, random fragments, and tiny ideas that technically existed but still needed to be placed somewhere. And sorting them by hand was annoying in a very specific way.

Every note asked for a decision: is this a project, an area, a resource, a daily note, or just trash with good marketing?

That was the real trigger for the first version of Friday.

Not “let’s build an AI friend.” More like: can I stop manually filing notes like an underpaid intern in my own brain?

Friday reporting: This is probably the first honest product insight here: the assistant was born from a boring operational pain, not from a fantasy of companionship. Boring pains are usually better product seeds.

I started iterating on an assistant that could understand the vault structure, move notes, patch daily notes, search context, and eventually become reachable from Telegram too. Telegram was still important, but at that stage it was the capture surface, not the assistant itself.

That distinction matters. Telegram is a great input pipe. It is not a memory system.

Streams are good for capture. They are terrible for memory. Stuff disappears. Context gets buried. There is no real structure. If everything lives in chat, your “second brain” becomes a very long scroll of unresolved intentions. Which is exactly the kind of thing AI should help reduce, not make worse.

Why Obsidian stayed in the center

The reason Obsidian works for this is almost boring: everything is just Markdown files. No proprietary magic. No database I can’t touch without 16-symbol passwords. No SaaS wall. My notes are readable by me, by scripts, and by agents.

That matters a lot. If I want an AI system that actually understands work and thinking, it needs access to the place where work and thinking already live.

So Obsidian became the memory layer.

Daily notes capture what happened. Project notes track what I’m building. Area notes hold long-term responsibilities. Resources store useful references. It’s not perfect, and I’m not trying to build some sacred productivity cathedral here, but it gives the assistant something much better than chat history: a structured environment.

Friday reporting: I would frame Obsidian less as “where notes live” and more as the filesystem for personal context. The important thing is not that it stores Markdown. The important thing is that the agent can operate inside the same structure you use to think.

When the workflow flipped

Before Friday understood Obsidian, it was mostly reactive. Useful, but still kind of shallow. Once it could work with the vault, it started becoming contextual. It could look through old notes. It could connect today’s question to something I wrote two weeks ago. It could help write a review and pull in details from daily notes I didn’t even remember writing.

That was the first moment where it stopped feeling like “prompting” and started feeling like an actual thinking partner.

At the start, Obsidian was still the main interface. I opened the vault, searched manually, moved notes around, looked through folders, and used Friday more like a copilot sitting next to the system. It helped, but I was still the one operating Obsidian directly most of the time.

Then at some point the workflow flipped upside down.

I started using Friday more than I used Obsidian itself. Obsidian stayed important, but more as the underlying knowledge base than as the main interface. Now I mostly open Obsidian to read notes or check the final result. I don’t really search through Obsidian manually. I don’t rely on graph views, tags, dashboards, or any fancy productivity decoration. Most of that work moved into Friday.

Friday became the wingman.

That word fits better than “operator” or “interface.” A wingman watches your wing. They know your patterns. They notice when something is off. They are useful because you trust them with context, not because they have a shiny dashboard.

Friday learned the shape of the vault, my PARA structure, my daily notes, my habits, my style of asking, and the kinds of answers that are actually useful for me. That changed the feeling of the whole system. Instead of me adapting myself to the tool, the tool started adapting to how I already behave.

Friday reporting: This is probably the strongest part of the story. The win was not “AI inside Obsidian.” The win was that Obsidian became infrastructure, and I became the trusted layer watching your wing.

The wow moment

The moment when I became properly bullish was watching the system answer a question that would have been annoying for a human version of me to answer.

I asked something like: “Friday, when was the last time I connected with this friend?”

And in a matter of minutes, it traced the person through my CRM, found the relevant notes, connected them with older daily reflections where I described the experience, and gave me the answer. Not a generic “you should reconnect with friends” answer. The actual context. The actual trail up to the place and type of beer we had been drinking. The kind of thing that technically existed somewhere in my system, but would have taken enough searching that I probably would not have done it manually.

That was the wow moment.

Part of it was personal: seeing other people react with genuine amazement when I casually asked Friday something and it pulled the answer out of my own messy life infrastructure. But the deeper part was realizing that this is what personal AI is supposed to feel like.

Not a chatbot performing intelligence. A system that can recover context from your life faster than you can.

Current setup

The stack kept evolving from there.

Figure 3. Current setup: a VPS shard for lightweight Telegram work that runs 24/7, and a MacBook shard for heavy lifting with the extended toolset. Both use Pi coding agent, powered by a single $20 Codex subscription.

Telegram became the mobile command surface. Obsidian became the long-term memory. Local tools let the assistant read files, edit notes, create reminders, check calendars, search the web, transcribe voice messages, generate PDFs, and run small workflows. Skills turned repeated workflows into reusable procedures: morning briefs, evening reviews, finance reviews, vault search, transcription, research, planning.

The interesting part is not that any of these tools are individually impressive. Most of them are pretty normal. The interesting part is what happens when they are connected.

What Friday can do now

Right now Friday can do a pretty wide range of small-but-real jobs:

• read, edit, sort, and patch notes in my vault
• update daily notes and connect old context
• create reminders and inspect my calendar
• transcribe voice notes
• summarize Telegram threads
• do web research and extract articles
• draft messages and structured writing
• generate PDFs
• help with finance reviews
• run morning and evening reviews
• turn messy thoughts into something usable

It is not one huge “AI app.” It is more like a bunch of sharp tools behind one interface.

Friday reporting: The capability list sounds random until you see the pattern: capture, retrieve, decide, execute. That is the loop. Everything else is just a tool plugged into one of those verbs.

I can send a messy voice note from Telegram and have it transcribed into the right daily note. I can ask “what am I missing?” and get an answer grounded in my actual recent reflections, not generic self-help fog. I can ask for a research brief and get something that understands my current projects. I can ask it to draft, summarize, plan, remind, search, reorganize, or turn a half-formed idea into a first version of an article like this one.

That wingman is Friday.

Not a chatbot. Not just a local operator for my personal system. More like something that watches my wing across notes, people, projects, habits, and decisions.

It’s still not the final version. I don’t want to oversell it as some perfect always-on chief of staff. It breaks. It needs better workflows. Some parts are still duct tape. Some parts are very cool duct tape, but duct tape nonetheless.

But the direction feels right.

What I learned

If I had to pull out the lessons for someone who wants to build an AI agent for themselves, I would start with three.

1. Context matters more than almost anything else

A personal assistant gets better when it knows more about your projects, habits, notes, writing style, current problems, and the weird background assumptions you do not want to explain every time. The more useful context it has, the more often it can guess what you actually mean instead of answering the literal sentence you typed.

And if you are not the kind of person who likes to write down observations, you would need to become one in order to leverage this kind of technology.

But I would be careful with the lazy version of this insight. It is not just “dump more data into the model.” Raw context becomes noise very quickly. The real leverage is giving the assistant context in a form it can use: structured notes, clear folders, daily logs, recurring workflows, and enough examples of what good output looks like.

Friday reporting: More context helps only if it increases judgment. Otherwise you are just building a very expensive hoarder.

2. Give the system time to fit you

It will not feel like a real personal assistant on day one. Probably not in a week. Maybe not even in a month. The useful version appears after enough cycles: you ask, it answers, you correct it, it learns your patterns, you adjust the workflow, and slowly the system stops feeling generic.

That requires some commitment. Not in a romantic “AI companion” way. More like training a good operator. If you never give feedback, never write things down, never let it touch real workflows, and never correct it when it misses, it will stay a chatbot with a cute face.

3. Build around repeated behavior, not vibes

The assistant became useful because it attached to things I already do all the time: capturing thoughts, sorting notes, checking priorities, drafting text, reviewing days, researching topics, creating reminders. Those loops existed before Friday. Friday just made them faster and less annoying.

That is why I think the wrong question is “what personality should my AI assistant have?” Personality is nice, but it is downstream. The better question is: where does my life already create friction every day, and what would it look like if an agent could quietly remove 30% of it?

Where to start

If someone asked me how to start building something similar, I would not start with the model or the framework. I would start with the places where their life already leaks context.

For one person, that might be notes scattered between Notion, Apple Notes, and Slack. For another, it might be calendar chaos, half-remembered conversations, unread PDFs, forgotten follow-ups, or decisions buried in chat history. The exact tools do not matter that much. The pattern is the same: useful context exists somewhere, but it is not available at the moment when it would change the answer.

That is where a personal AI system becomes interesting. Not when it writes a better generic email, but when it can connect your own inputs, history, habits, and workflows into something you can actually use.

That is the result I care about with Friday: not that I have a cool AI assistant, but that more of my own context is available to her at the moment when I need it.

written by Yarema Kertytsky with Friday, Lviv, Ukraine, June 2026.