Pick from 43 personality-matched tuning files — or take a 5-min test to find yours. Paste into Claude, ChatGPT, or Gemini and superpower your AI agent.
You spend five minutes once. After that, every conversation with your AI starts where you'd otherwise have to drag it to manually.
Five research-grade instruments to choose from: MBTI, Big Five, Enneagram, DISC, or Attachment. No signup, no email, no progress bar disguised as a sales funnel.
I'd rather have the underlying model than the bullet list.
Your result compiles into a short Markdown file — behavior rules for your agent, not a horoscope. A dozen lines that cancel the model's worst defaults.
# INTJ — Architect Lead with the underlying model. Skip the encouraging preamble. - No "great question" openers - One structured answer, not a menu - Push back on flawed premises
One paste into your system prompt — Claude, ChatGPT, Cursor, Gemini, or any API. No SDK, no subscription, no lock-in. Works the next time you press send.
Same prompt, same model — only the tuning file changed. The left is what Claude says to everyone. The right is what it says once it knows you're an INTJ who wants the underlying model, not a bullet menu.
Great question! Here are a few approaches: 1. Break the project into smaller, more manageable chunks 2. Take a short break and come back with fresh eyes 3. Talk it through with someone to externalize the problem 4. Identify what specifically feels stuck Which resonates? Happy to go deeper on any of them.
What's the underlying model of the project — what depends on what, and where's the blocker? "Stuck" is usually a missing piece of the structure, not a missing piece of effort. If you can name the upstream dependency that hasn't resolved, the next move is obvious. If you can't, the work is to map it, not push through.
Tuning isn't about a personality test on the wall — it's about removing friction from every reply your agent gives.
Stop typing "be more direct," "skip the recap," "no caveats." The model already knows.
Cut 30–60% of the words you don't read anyway. Same depth, less noise.
No more bullet menus when you wanted a recommendation. No more recommendations when you wanted options.
If you're the type who wants challenge, you get challenge. Not a polite hedge wrapped in three disclaimers.
Use the same file in Claude, GPT, Cursor, your own API call. Switch models, keep the tuning.
Frontier models are trained on the population, then RLHF'd toward the safest, most agreeable response anyone could want. That's a sensible default for billions of users — and a mediocre one for any individual.
A personality test is the cheapest known instrument for telling a model how you differ from the average. AgentTune is just the bridge: the test gives you a type, the type compiles to a Markdown file, the file becomes context the next time you press send.
No magic. No new model. Just the right context, written down once.
One Markdown file lives in your password manager, your dotfiles, or a note. Drop it wherever you talk to a model.
Your AI pushes back instead of agreeing. Names flawed premises. Stops asking what kind of help you'd like — and just gives it.
No "would you like me to add tests?" No 80-line apology when the diff is wrong. Just the code you'd have written, faster.
It stops adding em-dashes everywhere and proposing structures you'd never use. It edits your draft instead of rewriting it.
All five tests return the same kind of file. Pick the framework you already believe in — or start with MBTI if you're undecided. You can stack files later.
No email, no account. Take a test, get a file, leave.
The library is open. Fork it, edit it, ship your own.
Tests run client-side. Results never leave your browser.
It's a Markdown file. Save it, version it, share it.
Five minutes for the test. Ten seconds to paste the file. From the next reply onward — every conversation, every model — your agent already knows the shape of you.
# INTJ — Architect Lead with the underlying model, then the conclusion. Skip the encouraging preamble. - No "great question" openers - One structured answer, not a menu of options - When uncertain, name it in one line and pick anyway - Push back on flawed premises directly