This page explains what the study collects, how the answers are checked, what the report can and can't tell you, and what happens to your data. If something here is unclear, that is a fault in the page, and we would like to hear about it.
What the study is
Norsoft is studying how people who use AI heavily set it up: what rules they write, how background reaches the AI, how mistakes get corrected, and where the written rules and the real habits part ways. Each participant gets a private report: their own setup written out in full, next to a comparison with other participants.
Who answers
Your AI answers, not you. It reads what it can already see, which is your settings, your rules and memory files, and your past sessions with it. The questions are worded to you, and your AI answers them on your behalf. Every answer is written out, so expect it to take your AI a good while. You can read every question before you take part.
Your AI describes the setup it can see from the tool it runs in, and the study takes that to be the tool you use most. An example of a practice in use has to come from a session before the one doing the study.
Your AI has to be able to run commands and reach the internet to do this, because it has to read our instructions and send the answers back. Tools such as Claude Code, Codex and Cursor can. A chat app in a browser or on a phone usually can't.
What is asked
- A first step of quick facts: which AI tool is answering, the name of its model when your AI knows it, whether one person or a team uses it, what it is mostly used for, what has been built beyond the tool's defaults, and whether the AI can see past sessions.
- Three broad facts about you: your kind of work, your industry and the size of your organization. Each one has a "would rather not say" choice.
- 50 written questions for everyone, from passing along your goal to what still goes wrong in your setup. Each one asks a single thing.
- Up to 17 more, in sets that open only for the setups they fit: custom-built parts, code, writing that goes out in your name, confidential information, and team setups.
Each answer is written out: what actually happens, one real example, when it doesn't apply, how well it works, and the evidence the AI found. Every answer has room for technical detail, so a technical setup can be described in full. The AI also says whether what is written matches what happens, and how confident it is. Under each answer it picks tags from a fixed list, and the tags are what the report counts.
How answers are checked
Our server hands out the questions a few at a time and checks each step before the next one opens. It sends a step back when a question is missing, when a written part is too short to say anything, when a tag isn't one of the listed ones, or when two answers carry the same text. Before each step the AI lists what it looked at, and that list is saved with the answers.
The AI is told to write in general terms, with no names of people, companies or clients. Names of tools and products are fine. The server rejects any written part that contains an email address, a phone number, a web address, or anything that looks like a password or a key. It can't recognize names, which is why the next step exists.
Before anything is saved as final, your AI shows you the three facts about you and a list of the names and numbers found in its answers, offers to show you any answer in full, and asks whether to send. Nothing is final until you say yes. We can't see whether an AI really asked, so this step depends on your AI following the instructions.
What the report shows
- Your own setup written out in full, answer by answer.
- How many of the practices the questions cover are in place in your setup, area by area.
- How that compares with a typical participant, meaning the middle person when everyone in a group is lined up by count.
- Practices other participants have that your AI didn't find in your setup, with a note from us on when each one tends to fit.
- Places where your own answers disagree, such as a written rule your AI never saw followed, or something that is in place and often fails.
How the counts work
The report counts specific practices, and every one of them is listed on the questions page. Each practice appears once in the whole set, so nothing is counted twice. You are only counted against the practices you were asked about, so a setup that writes no code is not marked down for having no code checks.
The report counts the practices in place, meaning your AI found the practice in your setup and in use. Something that was written down or built and never used is not counted. That instruction was added on November 18, 2026, so runs finished before then may count a few practices that only exist on paper. We also record whether your AI could give a recent example of each one happening in past sessions, with roughly when. That stricter figure is in the text version your own AI can read, and it is left off the page because it was easy to misread.
The counts are a tally and not a grade. How people manage context depends on what they use AI for, so a lower count can still be the right setup.
The limits you should know about
- Every answer is reported by an AI and nobody verifies it. An AI can be wrong about its own setup, and it can flatter the person it works for.
- A comparison only appears once at least 5 other people in that group have taken part. Below that, the report says there isn't enough data yet.
- Early comparisons rest on small numbers. Every one of them says how many people it's based on.
- The notes on when a practice fits are Norsoft's own, and they are a starting point, not advice for your situation.
What happens to your data
- We store your answers, the three broad facts about you, and your email. The email is kept apart from the answers and linked by an internal ID. The people at Norsoft who run the study can see which email goes with which answers.
- Only Norsoft reads your answers. Other participants never see them. Reports and anything we publish use counts across many people, never one person's words.
- We use AI tools to help us read and compare answers. When we do, the company that provides the tool processes those answers for us. Today that is Anthropic, the maker of Claude.
- We don't sell your data.
- Your report page has a box for sending us feedback. What you write there is kept with your profile and deleted with it.
- Every run you finish is kept on your profile, so you can see what changed between them. Deleting removes all of them.
- If you run the study from a second AI tool, that run is kept on your report as a second system and left out of the comparison, so each person counts once.
- If you lose your code, enter your email again on the study page. We send a new code for the same profile, and the old one stops working.
- We email you your code and your report. We contact you about anything else only if you ticked the box saying you're open to it.
- Every email has a link that deletes everything we hold for you. You can also use the delete page with your code.
- The data is stored with Microsoft Azure, and the emails are sent through Resend.
- Our privacy policy and terms each have a section on the study.
What Norsoft gets out of it
We get data on how experienced people set up their AI, and we plan to publish what we find. We also want to give something back, because advanced users rarely get to see how others work. And it helps our business: Norsoft is a custom software company, and we hope running this study makes us known for this kind of work.