People ask me what my everyday AI setup looks like, and I think they expect a list of eight apps and a subscription total with three decimal places. The real answer is less exciting and more useful. My setup is one main assistant, one second opinion, a notebook, and a short list of rules. I do not write code, I do not build automations, and I do not chase every new release. If you are learning how to use AI as a beginner, this is the setup I would hand you before you spend money on anything.
This is the sixth entry in The AI Test Log, and the first in the Workbench Skills column. The Test Bench posts document experiments. This column documents the habits and skills that make those experiments possible. If you are changing careers, picking up a new skill, or just trying to stop losing your afternoon to repetitive work, start here.
Why I Built a Setup Instead of Chasing Tools
The Month I Tried Everything
Early on, I did what most people do. I signed up for seven tools in two weeks. I had a writing assistant, a note-taking app with AI built in, a meeting summarizer, an image generator, a transcription service, and two things I cannot remember now. By the end of the month I was paying for six subscriptions and opening four of them, and none of my actual work had gotten faster.
The problem was not the tools. The problem was that I had no workflow for them. Every task started with a decision about which app to open, and decisions are expensive when you are tired.
The Question That Fixed It
I wrote one question on a sticky note and put it above my desk: what am I actually trying to finish?
That sounds obvious. It changed how I used everything. Instead of asking what a tool could do, I started with the task and worked backward. Most weeks, the answer was that I needed one assistant and a piece of paper.
The Setup I Actually Use
The Main Assistant
I use one general-purpose assistant as my default. It handles writing, structuring, explaining, and translating between jargon and plain English. It is the tool I open first for almost everything, and I have used it long enough to know its failure patterns.
Here is what I mean by failure patterns. It leads with the most common answer instead of the answer for my specific situation, which is what happened when I asked it to diagnose my trailer wiring. It also gets more confident as my questions get more specific, even when the underlying information gets thinner. Knowing those two habits has saved me more time than any feature.
The Second Opinion
I keep a second assistant for one purpose: checking the first one. When an answer matters, I paste the question into the second tool and compare. If both agree, I still verify anything factual. If they disagree, I know I am in territory that requires an outside source.
This is not a technical setup. It is a habit. Two tools cost me less than one mistake.
The Notebook and the Test Log
Every experiment goes in a paper notebook with the same nine fields: problem, setup, tools, test, what worked, what failed, what needed human judgment, cost and time, and final verdict. That structure is the single most valuable part of my setup, and it costs about four dollars.
Writing the verdict by hand forces me to commit. A tool I cannot recommend out loud is not a tool I should keep paying for.
What I Pay For
One main assistant subscription, one second assistant subscription, and nothing else dedicated to AI. I do not pay for a meeting summarizer because my meeting notes are short. I do not pay for an image tool because I do not need one weekly. I do not pay for a transcription service because I tested one and found I never used it.
The rule is simple. If I have not opened a tool in thirty days for real work, I cancel it.
How the Setup Works on a Real Task
Step One — Write the Problem Before Opening Anything
I write one sentence on paper: what I need to finish and what done looks like. That sentence goes into the assistant as the first line. It sounds small. It prevents the most common failure, which is asking a vague question and getting a generic answer that I then spend twenty minutes editing.
Step Two — Ask for Structure First, Content Second
For anything longer than a paragraph, I ask for an outline before I ask for prose. Outlines are cheaper to fix. If the structure is wrong, I have wasted ninety seconds instead of twenty minutes.
This is also how I get better output without learning prompt jargon. "Give me the structure first" is plain English, and it works.
Step Three — Ask Where the Answer Came From
For anything factual, I ask the assistant to say which part of my input it used and which part it inferred. That question surfaces the difference between reading and guessing. It is the single most useful sentence I have added to my workflow this year.
Step Four — Verify Outside the Chat
Anything that touches money, rules, safety, or a specification I cannot check gets verified somewhere else. A phone call, a government page, a manual, a supplier, or a person who does the work professionally. The assistant gets me to the question. It does not get to be the answer.
Step Five — Write the Verdict by Hand
After the task, I write two lines in the notebook. Did this save time, and would I do it again? That is it. No dashboard, no tracking app.
The Rules That Keep the Setup From Taking Over
I keep five rules taped inside the notebook cover.
No compliance questions. No allergen or ingredient statements. No pricing decisions. No legal language. No medical or safety specifications. Those five categories stay with a human who is accountable, and they are not negotiable.
I also keep one boundary about data. I do not paste customer names, account numbers, addresses, or anything I would not want forwarded. For most tasks, none of it is necessary, and it is a habit worth building before you need it.
What I Do Not Use It For
I do not use AI to write in my voice from scratch. I use it to tighten something I already wrote, because tone drifts when a model has no examples, and it still drifts when it has two.
I do not use it to decide. I use it to lay out options. Deciding is the part where my judgment is the product, and outsourcing it makes my work worse.
I do not use it to replace a professional. If a task requires a license, a certification, or a signature, it requires a person.
What Still Needs Human Judgment
Everything that depends on context I have not written down. How tired I am, how a customer is feeling, what my budget can absorb this month, and whether a plan is realistic for the time I actually have. No model knows those things unless I say them, and I usually do not say them because I am in a hurry.
The other thing that stays with me is the final read. I read everything out loud before it goes anywhere. If it sounds like a press release, it gets rewritten. That single check catches more problems than any prompt I have tried.
Cost and Time
Two subscriptions, both of which I use weekly for real work. Setup time was about two hours spread across a month, most of it spent deciding what to cancel. Time saved per week is harder to state honestly. On writing and structuring tasks, I would estimate three to four hours. On anything requiring verification, the savings are smaller, because verification is the work.
Final Verdict
Keep it.
This setup is boring, cheap, and portable, and it has survived six months of real use across a consulting practice, a sauce business, a trailer restoration, and a garden. It does not require coding. It does not require a course. It requires one assistant, one second opinion, a notebook, and the discipline to verify before you trust.
If you are starting from zero, copy the nine-field test format first and add tools later. The format is what makes the tools useful.
Take it apart first. Then ask AI.
No comments yet — be the first to share a thought.