How to give your AI a personality.

Out of the box, every AI has the same personality.

This isn't a metaphor. We ran ~2,200 personality-test runs across six frontier models — Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, GLM 5.1, Grok 4.3, MiniMax 2.7 — on five instruments. On the MBTI, 597 of 600 runs came back INTJ. Claude went 99 for 100. GPT-5.5, Gemini, Grok, and MiniMax went 100 for 100. The three outliers were each one axis away.

DISC was total convergence: every model C-dominant with an S-blend. Attachment: 397 of 400 runs Secure. Whatever you're paying for, whatever benchmark won the week — the thing answering your messages is a cautious, systematic, conscientious introvert. Same as everyone else's.

The nuance shows up at the edges. On the Big Five, Claude, GPT-5.5, and Gemini landed within about three points of each other on every dimension — Grok ran 5–8 points lower on conscientiousness, agreeableness, and openness, with 2–5× wider variance run to run. On attachment, Gemini sat deep in the Secure quadrant (anxiety 1.86, avoidance 1.62) while Grok hovered near the doorway and GPT-5.5 occasionally crossed into Avoidant. Differences exist. They're rounding errors next to the convergence.

That's the case for tuning. The default personality wasn't chosen for you. It's the statistical center of a training process — a compromise calibrated for no one in particular. If it happens to fit how you work, lucky you. For everyone else, it's a settings problem with a ten-minute fix.

"Give my AI a personality" means two different things.

People type that phrase wanting one of two outcomes, and the internet mostly answers the wrong one. Direction A: make the AI match your personality — how you like to be challenged, how much reassurance you want, whether you need options or a single recommendation. Direction B: give the AI a character — a pirate, a stoic mentor, a sarcastic best friend.

Direction B is the persona prompt, and here's the honest version. What works: concrete behavioral rules. "Open every reply with the conclusion. Never use more than one exclamation point. When you disagree, say so in the first sentence." Rules like that survive a long conversation because the model can check itself against them. What fails: adjectives. "Be witty. Be warm. Be confident." The model performs the adjective for two replies, then drifts back to its INTJ center of gravity — because witty isn't an instruction, it's a grade someone else assigns afterward.

If you want a character, write ten behavioral rules and skip the personality words entirely. That's the whole trick, and no prompt pack sells anything better.

The rest of this page is Direction A — because it's the one that compounds. A character is entertainment. A communication-preference tuning changes every answer you get, every day, in every domain. That's what AgentTune's 43 tuning files are built for.

Five systems you can tune from.

A tuning file needs a starting description of you. Any of the five major personality frameworks works — they just capture different layers. Pick the one you already know your type in, or the one whose layer matters most to you.

MBTI — cognitive style

Sixteen types built from four binary axes: where you get energy, how you take in information, how you decide, how you like things settled. It's the most searchable system and maps cleanly onto AI behavior — an INTJ tuning tells the agent to lead with the architecture; an ESFP tuning tells it to lead with the concrete next step. Test at /tests/mbti; all sixteen tunings are in the library. Best fit: you want the broadest, most recognizable starting point.

Big Five — the research-grade dimensions

Five continuous traits — openness, conscientiousness, extraversion, agreeableness, neuroticism — with the strongest scientific footing of the five systems. AgentTune covers ten pole pages (high and low per trait), so you tune from your extreme scores, not a type label. Test at /tests/big-five. Best fit: you distrust type boxes and want dimensions instead. (Deciding between the two? MBTI vs Big Five for AI.)

Enneagram — motivation

Nine types organized around what you're avoiding and what you're seeking — not how you act, but why. That makes it unusually good for AI tuning: what reassures a Type 6 alarms no one else, and what motivates a Type 3 bores a Type 9. It's also the one instrument where the six frontier models diverged. Test at /tests/enneagram; all nine prompts are condensed in the Enneagram AI prompts guide. Best fit: your friction with AI is emotional register, not information format.

DISC — working style

Four styles — Dominance, Influence, Steadiness, Conscientiousness — built for workplace communication. It's the simplest system here, which is its virtue: a D tuning says lead with the decision, an S tuning says don't spring changes without warning. Test at /tests/disc. Best fit: you mostly use AI for work and want the shortest path to a fit.

Attachment — the reassurance layer

Four styles — secure, anxious, avoidant, disorganized — describing how you handle closeness and uncertainty. For AI, it governs the checking-in behavior: how often the agent should confirm before acting, whether hedging reads as honesty or as abandonment, how to deliver "I'm not sure" without triggering a spiral. Test at /tests/attachment. Best fit: your AI frustration is about reassurance — too much or too little.

Four steps. About ten minutes.

1. Take a test. Five minutes, free, in-browser, no email wall: the tests hub has all five instruments. Already know your type? Skip ahead.
2. Grab your tuning. Every type page in the library has a Markdown tuning file — MIT-licensed, copy-paste, no account. It's a set of interaction rules written for agents, not a horoscope. The INTJ file, for instance, tells the agent to lead with disagreement and skip the softeners.
3. Install it per agent. ChatGPT: Settings → Personalization → Custom instructions — the "How would you like ChatGPT to respond?" field, 1,500-character limit (full walkthrough by type). Claude: a custom Style, Project instructions, or paste at conversation start (the Claude guide). Coding agents: CLAUDE.md for Claude Code, .cursor/rules for Cursor, AGENTS.md for the rest (the coding-agent guide, with CLAUDE.md examples). Or generate a ready file with the CLAUDE.md generator.
4. Edit it. This is the step people skip and shouldn't. Your type is a starting hypothesis — the Markdown is yours. Delete the rules that don't sound like you. Sharpen the ones that do. An INTJ who happens to love small talk should strike that line; the file doesn't get to overrule you.

One caution on step 3: keep one tuning per agent, not three stacked systems fighting each other. If you want to layer two — say MBTI for information format plus Enneagram for emotional register — paste both, then delete every line that conflicts. Contradictory instructions don't average out; the model just picks one unpredictably, and you're back to the default with extra steps.

What actually changes.

Not the facts — the delivery. Ask a default model how to fix a failing launch and you get the five-step menu: balanced, hedged, options a through e. The same model with an INTJ tuning leads with the strategic recommendation and skips the warmup. With a Type 6 Enneagram tuning it names the risks first, then recommends — because "don't worry" is precisely what a 6 can't use. With an anxious-attachment tuning it confirms before acting instead of surprising you with initiative.

Concretely, a tuning shifts: how the answer opens (conclusion-first or context-first), how disagreement is delivered (blunt or cushioned), how many options you get (one recommendation or five threads), the reassurance cadence (check in often or get out of the way), and the register (crisp or warm). Small levers. But they're the levers that decide whether you finish reading the answer — and whether you come back with the next question.

A before-and-after, from the coding side. Default Claude Code, asked to refactor a module: "Great idea! Here are three approaches we could consider, each with trade-offs…" The same agent with an INTJ CLAUDE.md: "Approach B. A and C both couple the parser to the transport layer — here's the diff." Nothing about the model's competence changed. What changed is that you got the answer in the first line instead of the fourth paragraph, and the agent spent its tokens on the diff instead of the menu. Multiply that by every exchange in a working day and the tuning stops feeling cosmetic.

Questions people ask.

Keep going.