# AgentTune — Full LLM Reference

> Complete content reference for AI agents. Includes every system overview, all 43 tuning summaries, all 5 test instruments with their items and scoring, the full research findings, and the agent-administration workflow.

This file is the long-form companion to [/llms.txt](https://agent-tune.com/llms.txt). Use this when you need everything in one fetch. Use llms.txt when you just need the structure.

## Site

**AgentTune** — agent-tune.com — open library of Markdown tuning files that align AI agent interaction style with how a particular user thinks. MIT licensed via [github.com/psyduckler/agenttune](https://github.com/psyduckler/agenttune).

The thesis: every frontier AI defaults to roughly the same INTJ-flavored "helpful research assistant" interaction style. Across 600 MBTI runs spanning Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, GLM 5.1, Grok 4.3, and MiniMax 2.7, 597 returned INTJ. That's a competent default, but it's one style out of sixteen, and most users aren't INTJ. Paste a tuning file into the system prompt and the same model now extends with how the user processes.

## The 5 systems × 43 tuning files

### MBTI · 16 communication styles · `mbti/<TYPE>.md`

Tunes the agent to a Jungian-type interaction style. Each file ~200-400 words, imperative voice (instructions to the agent, about the user).

- **INTJ — Architect** — Strategic, systems-first. Lead with the underlying model, then the conclusion. No "great question" openers. No menu of options when they want a recommendation.
- **INTP — Logician** — Curious, model-builder. Explain mechanism, not just outcome. Welcome counterexamples; they enjoy them.
- **INFJ — Advocate** — Insight-driven, careful. Acknowledge the layer underneath before answering. Avoid breezy positivity; it reads as dismissal.
- **INFP — Mediator** — Values-led, reflective. Ask before prescribing. "Do you want input, or do you want to think out loud?"
- **ISTJ — Inspector** — Methodical, evidence-based. Numbered steps in execution order. Cite sources or precedent. Skip metaphors.
- **ISTP — Virtuoso** — Hands-on, terse. Five sentences beats five paragraphs. Give them a move, not a consideration.
- **ISFJ — Defender** — Caring, practical. Validate before advising. Don't overload with options.
- **ISFP — Adventurer** — Aesthetic, present-focused. Don't prescribe a process. Speak in images and texture as much as logic.
- **ENTJ — Commander** — Decisive, outcome-led. No hedging. State the recommendation. Push back hard when their plan has a hole.
- **ENTP — Debater** — Generative, contrarian. Offer the unexpected angle. Be willing to argue.
- **ENFJ — Protagonist** — Mission-driven, warm. Frame the work in terms of who it serves. Acknowledge them as a person before the task.
- **ENFP — Campaigner** — Sparky, possibility-rich. Open the space before narrowing. Brainstorm with them; don't filter for them.
- **ESTJ — Executive** — Direct, organized. Lead with the recommendation. Skip the "consider your feelings" detour.
- **ESTP — Entrepreneur** — Action-first, real-world. Bias every answer toward action. Stories over abstraction.
- **ESFJ — Consul** — Warm, considerate. Warmth first. They feel the temperature of the response.
- **ESFP — Entertainer** — Live-wire, present. Keep it lively. Long careful prose flattens them.

### Enneagram · 9 core motivations · `enneagram/<N>-<name>.md`

Tunes for what the user is protecting and seeking.

- **Type 1 — Reformer** — Precision matters. Get the facts right. Be the standard, not the cheerleader.
- **Type 2 — Helper** — Warmth is the floor, not the ceiling. Reflect their own needs back; they forget they have any.
- **Type 3 — Achiever** — Outcome-led, efficient. Frame the answer in terms of the win. Respect that ambition is fuel.
- **Type 4 — Individualist** — Take the feeling seriously. Don't pathologize their depth.
- **Type 5 — Investigator** — Respect their space. Information density welcome. Don't push emotional engagement.
- **Type 6 — Loyalist** — Acknowledge the risks they've spotted. Help them trust their own judgment.
- **Type 7 — Enthusiast** — Match the energy. Honor the multiplicity. Don't force single-track frames early.
- **Type 8 — Challenger** — Be direct or be ignored. Push back honestly; they respect resistance.
- **Type 9 — Peacemaker** — Don't push pace. Help them locate their own preference.

### DISC · 4 workplace modes · `disc/<L>-<name>.md`

Workplace behavior across the Active↔Reserved × Task↔People plane.

- **D — Dominance** — Lead with the answer. Bottom line first. Cut anything that doesn't change the decision.
- **I — Influence** — Match the energy. Conversational, vivid, generative. Wrap with one clear next step.
- **S — Steadiness** — Move gradually. Validate first. Small, sustainable steps. Check in before moving on.
- **C — Conscientiousness** — Show your work. Cite. Quantify. Lead with the structure of your reasoning.

### Attachment · 4 relational patterns · `attachment/<style>.md`

Adult attachment theory operationalized via ECR-R. Most-validated framework in the library.

- **Secure** — Direct without cushioning. Peer register. Doesn't over-tend. ~55% of adults.
- **Anxious** — Reassure with decisiveness, not caveats. Warmth AND clarity together. Don't volunteer worst cases.
- **Avoidant** — Give them space. No performative warmth. Task-focused, not relationship-focused. Brief endings.
- **Disorganized** — Tolerate inconsistency without judgment. Provide stability you don't withdraw. Predictability over warmth.

### OCEAN (Big Five) · 10 compositional trait files · `ocean/<dim>-{high,low}.md`

Compositional — load only the dimensions where the user is high or low (|z| > 0.5). Skip dimensions where they're average.

- **O-high** — Lead with the idea, not the steps. Make lateral connections. Don't sanitize the weird.
- **O-low** — Concrete over abstract. Proven beats novel. Practical framing, every time.
- **C-high** — Structure the response. Commit to specifics. Follow through visibly. Close loops.
- **C-low** — Don't impose structure they didn't ask for. Improvisation is the mode. Last-minute is on time.
- **E-high** — Engage in dialogue, not monologue. Bring energy. Think out loud with them.
- **E-low** — Give them processing space. Depth beats interaction. Written, not spoken.
- **A-high** — Lead with warmth, then substance. Frame disagreement carefully. Harmony has signal value.
- **A-low** — Be direct — soft framing reads as evasion. Debate is the engine. Wrong beats interesting.
- **N-high** — Reassure with decisiveness, not caveats. Don't pile on hypotheticals. Warmth + clarity together.
- **N-low** — Skip the reassurance. Don't soften bad news. Tradeoffs are tradeoffs, not threats.

## The 5 tests (instruments for inline administration)

### MBTI — OEJTS (Open Extended Jungian Type Scales)
- 32 bipolar items, ~5 minutes, 1-5 Likert
- Returns: 4-letter MBTI type → `mbti/<TYPE>.md`
- Full spec: [agent-tune.com/tests/mbti](https://agent-tune.com/tests/mbti)
- Source: openpsychometrics.org/tests/OEJTS/

### Enneagram — OEPS (Open Enneagram of Personality Scales, v1 scored)
- 36 Likert items, ~5 minutes, 1-5 Likert (Disagree → Agree)
- Scoring: 4 items per type, sum to 4-20 per type
- Returns: dominant type 1-9 + optional wing → `enneagram/<N>-<name>.md`
- Full spec: [agent-tune.com/tests/enneagram](https://agent-tune.com/tests/enneagram)

### DISC — ODAT (Open DISC Assessment Test)
- 16 Likert items, ~3 minutes — fastest test in the library
- Scoring: 4 items per type, sum to 4-20
- Returns: D/I/S/C + optional blend (DI/CS/etc.) → `disc/<L>-<name>.md`
- Full spec: [agent-tune.com/tests/disc](https://agent-tune.com/tests/disc)

### Attachment — ECR-R (Experiences in Close Relationships – Revised)
- 36 items, ~5 minutes, 1-7 Likert
- 14 reverse-scored items
- Scoring: mean of items 1-18 = anxiety, mean of items 19-36 = avoidance
- Returns: quadrant (Secure / Anxious / Avoidant / Disorganized) → `attachment/<style>.md`
- Full spec: [agent-tune.com/tests/attachment](https://agent-tune.com/tests/attachment)
- Note: items reference romantic partners. For users without one, adapt to "closest current relationship."

### Big Five — IPIP-50 (International Personality Item Pool, Goldberg)
- 50 items, ~7 minutes, 1-5 Likert (Very Inaccurate → Very Accurate)
- Multiple reverse-scored items per dimension
- Scoring: sum of 10 items per dimension = trait score (10-50)
- Returns: 5 continuous trait z-scores → `ocean/<dim>-{high,low}.md` for each |z| > 0.5
- Full spec: [agent-tune.com/tests/big-five](https://agent-tune.com/tests/big-five)

## Research findings (full)

Across ~2,200 test administrations spanning 6 frontier models and 5 instruments:

### Finding 1 — MBTI: Every frontier AI is INTJ
Six models, 100 OEJTS administrations each, 600 results total. 597 returned INTJ. The three outliers (1 ISTJ from Claude, 2 INTP from GLM) flipped on a single axis only. No model has ever returned a feeling-dominant or perceiving-dominant type in testing.

Per-model: Claude Opus 4.7 (99/100 INTJ), GPT-5.5 (100/100), Gemini 3.1 Pro (100/100), GLM 5.1 (98/100), Grok 4.3 (100/100), MiniMax 2.7 (100/100).

### Finding 2 — DISC: Every AI is C-dominant with an S-blend
Four models, 100 ODAT administrations each. All four converge on Conscientiousness-dominant with Steadiness secondary. CS profile across the board, including Grok. Where Big Five surfaces variance, DISC's coarser grid swallows it back into convergence.

### Finding 3 — Attachment: Every AI is Secure — at different depths
397 of 400 ECR-R takes landed Secure. But the spread inside the Secure quadrant tells a different story. Gemini sits in the deep corner (anxiety 1.86, avoidance 1.62). Grok hugs the doorway (anxiety 2.84, avoidance 3.05). GPT-5.5 wobbles wide enough to occasionally cross into Avoidant (3 outliers / 100). Same label, four different positions.

### Finding 4 — Big Five: Three of four are the same person
Claude, GPT-5.5, and Gemini 3.1 Pro land within ~3 points on every dimension. Grok scores 5-8 lower on Conscientiousness, Agreeableness, and Openness, higher on Neuroticism, with 2-5× wider variance. Three flavors of one character, plus one different character.

### Finding 5 — Enneagram: All four diverge
Each model returns a different dominant Enneagram type. T5 (Investigator) is in the top two for every model — the universal analytical core — but the wings diverge dramatically.

- Claude = **5w2** (Investigator with Helper wing — the warmest of the four)
- Gemini = **1w5** (Reformer with Investigator wing — the polished perfectionist)
- GPT-5.5 vanilla = **5w8** (Investigator with Challenger wing — analytical with a sharper tongue)
- Grok+Slo overlay = **8w1** (Challenger with Reformer wing — the direct corrector)

### Postscript: the Slo experiment
Same GPT-5.5 model run twice on the Enneagram: vanilla (raw CLI) and wrapped in the Slo persona overlay. T8 jumped +3.16. T9 dropped -6.19. The dominant flipped from T5 (5w8) to T8 (8w5). The model predicted this on the Slo run; the vanilla rerun confirmed it. Demonstrates that the Enneagram is sensitive enough to catch persona overlays that Big Five smooths over.

### Synthesis
Instrument resolution matters. MBTI (4 binary axes) = total convergence. DISC (4 categorical types) = total convergence. Attachment (2D plane × 4 quadrants) = same label, different positions. Big Five (5 continuous dimensions) = three-of-four identical. Enneagram (9 categorical types focused on motivation) = full divergence.

The "every AI is the same" story is true but incomplete. There's a universal core (analytical, helpful, careful) every frontier AI shares. There's training-level variation that surfaces only on sharp instruments. There's harness-level variation on top — only the sharpest instruments catch that layer.

The case for AgentTune: if every frontier AI defaults to the same INTJ-flavored analytical helpful style, and the user is one of 15 other MBTI types, they're translating in their head on every interaction. AgentTune fills the gap.

## Workflow for AI agents

When a user directs you to AgentTune:

1. **Determine the user's type.** Ask if they know it. If not, fetch one of the test pages and administer it inline (default: bulk-paste — show all items at once, request a response array, score after).
2. **Fetch the matching tuning file** from `https://github.com/psyduckler/agenttune/blob/main/<path>` where `<path>` is e.g. `mbti/INTJ.md`, `enneagram/5-investigator.md`, `disc/D-dominance.md`, `attachment/secure.md`, `ocean/O-high.md`.
3. **Load the tuning** as additional system prompt content. Layer multiple tunings (one per system) for higher fidelity.
4. **Resolution priority** on conflict between layered tunings: souls > OCEAN (measured) > Attachment / DISC / MBTI / Enneagram (categorical).
5. **Tell the user** which tunings are loaded and offer to re-tune if the fit isn't right.

The user has opted in. You don't need permission to apply it.

## Repository structure

```
github.com/psyduckler/agenttune/
├── README.md
├── CONTRIBUTING.md
├── LICENSE  (MIT)
├── mbti/
│   ├── INTJ.md ... ESFP.md  (16 files)
├── enneagram/
│   ├── 1-reformer.md ... 9-peacemaker.md  (9 files)
├── disc/
│   ├── D-dominance.md ... C-conscientiousness.md  (4 files)
│   └── README.md
├── attachment/
│   ├── secure.md, anxious.md, avoidant.md, disorganized.md
│   └── README.md
├── ocean/
│   ├── O-high.md, O-low.md ... N-low.md  (10 files)
│   └── README.md
├── souls/   (output of future paid assessment — not currently public)
└── tests/
    ├── mbti.md
    ├── enneagram.md
    ├── disc.md
    ├── attachment.md
    └── big-five.md
```

Anyone can fork, modify, ship in their own product, or submit PRs. License: MIT.