Build Your Personal AI System: Where to Start

Andrew Maryasov, AI consultant, founder of Auspex and Grow2.ai. A personal AI system starts not with tools but with one value loop: a context layer about you, one repeatable task delegated to an agent, and a log that separates “saves time” from “feels like it saves time”. Below — the assembly order, a four-week start, the traps everyone hits, and the honest line where the map ends.

TL;DR (30 seconds)

• Don’t copy someone else’s stack. A stack is the finale of someone else’s story, built for their tasks and budget; a copy gives you a collection of subscriptions, not a system.
• Context beats prompts. 90% of the result comes from files about you, your work and your goals; the perfect instructions are already built into the tools.
• Don’t trust your feelings at the start. In the METR study, developers felt +24% faster while measurement showed −19%. Only the log.
• First month = four steps: a context folder → a week of logging → instructions for anything done twice → one encoded judgment call → measure and kill.
• Budget drifts quietly: five dollars a day, then eighty a month later. Keep a daily cap — mine is $15.

Why not start by copying a stack

Starting by copying someone else’s stack is the most natural and the worst option. A stack is the finale of someone else’s story: assembled for their tasks, their budget and their head. Copy the finale and you get a collection of subscriptions plus the feeling that “something is off”.

Individual posts don’t help here: each shows one slice of an already finished system — a stack here, memory layers there, six-month numbers somewhere else. But “where to start” is a question about assembly order, and you can’t see it in the final slices.

The route: what to read and in what order

First — the principle without which nothing else makes sense: a system starts not with tools or prompts but with context. What the AI knows about you, your work and your goals is the value; chasing new tools gives only the illusion of progress, while results come from integrating a few — not collecting ten.

Next — a living build with honest numbers: six months of my personal AI operating system. What worked, where I caught myself doing silly things, and the beginner’s main trap — in the METR study developers felt +24% faster while measurement showed −19%. Forty-three percentage points between “feels” and “is”.

And as a reference for a full build — my AI stack 2026: about $300 a month, a council of three models, a few dozen specialized skills. Look at it not as a shopping list but as an example of what a system grows into over six months — if it grows out of context, not subscriptions.

The first month: four steps

Week one: one tool and a context folder. Five to ten plain-text files: who you are, what you do, your goals, your current projects. My main operating file is 40–60 lines, no more. And all week, run one repeatable task through the agent while keeping a log: what actually saved time versus what merely felt productive.

Week two: turn anything done twice into an instruction. If you repeated a task twice with the agent — the third time it should do it alone.

Week three: one repeatable judgment call — email triage, lead scoring, task prioritization — written down as rules. The act of writing the rubric will show how consistent your decisions have been so far.

Week four: measure and kill. Time, quality, cost. Anything that looks good on feelings and fails on numbers — throw away without regret.

The traps everyone hits

I did too. Editing AI drafts in areas where you are not an expert makes quality worse, not better. Inflating prompts instead of moving knowledge into files — prompts should get shorter while files grow; a prompt that went from 50 lines to 500 signals missing context, not missing cleverness. And after a third failure, don’t rephrase the request — add context. The error is almost always there, not in the model.

On money, separately: keep a daily limit. Costs drift quietly — five dollars a day, then eighty a month later, and in the moment it always feels “cheap”. My ceiling is $15 a day, and I don’t let myself cross it.

Where the map ends

The honest limit of this route: the map is generic, but everyone’s first value loop is their own. Which process to automate first, what exactly to put into the context files of your work, where your “the agent may” boundary lies — no article can tell you that.

Two paths

Yourself: four weeks by the plan above. Materials, order, traps — it’s all in the route; the rest of the blog covers individual layers, from the stack to the six-month experience.

With me: if you want to move faster and straight onto your own work — an advisory session on your situation: what to automate first and what your first value loop will consist of.

FAQ

Which tool should a personal AI system start with?

One — any modern assistant of Claude’s level. The question is not the tool but the context: five to ten plain-text files about you, your work and your goals will give you more than a second and third subscription.

How much does a personal AI system cost per month?

The start is the price of one subscription ($20–100). My system grew to ~$300 a month over half a year, but that is the finale of the story, not the entry ticket. What matters more is a daily spending cap: the drift from $5 to $80 a day happens unnoticed.

Why can’t you trust the feeling of productivity?

In the 2026 METR study, experienced developers using AI self-reported +24% speed while measurement showed a 19% slowdown. For the first month, keep a log: what actually saved time versus what merely felt productive.

What to do when the agent fails a task for the third time?

Don’t rephrase the request — add context. The error is almost always in missing materials, not in the model. A third failure is a zoom-out signal: what does the agent not know about your task?