Generic AI is predictableMake it personal

A system prompt — a Markdown file — so precisely calibrated to your personality, communication style, and values that any AI model running it produces outputs that sound authentically like you, not like a generic assistant.

It captures everything from your sentence cadence and vocabulary to your moral positions, professional persona, and hard limits. When deployed, the AI in that conversation stops performing helpfulness and starts performing you.

The methodology is grounded in over four decades of research across psycholinguistics, personality science, computational linguistics, and AI persona simulation.

The strongest external validation comes from two 2024–2025 studies: Park et al. (Stanford / Google DeepMind) demonstrated that LLM-based digital twins replicate survey responses with 85% accuracy compared to human test-retest reliability. Toubia et al. independently replicated this finding at 81.72% using a structured questionnaire method — the closest published architecture to ours.

Individual questionnaire questions map directly to validated instruments including the Big Five Inventory (BFI-2), Schwartz's Universal Values framework, and Pennebaker's LIWC programme. The Security & Integrity Layer is built to the OWASP Top 10 LLM Security standards (2025).

The table below is a transparent source list of the frameworks and studies used as references while building the pipeline.

This is a summary reference table intended for quick review. It focuses on the source itself and the publication details so you can scan the evidence base efficiently.

Deploy it in Claude Projects, a Custom GPT, Grok, OpenClaw, or any local model running via Ollama + Open WebUI or LM Studio. The implementation guide included with your delivery walks through each platform step by step.

The simplest starting point for most users is Claude Projects - paste the prompt into a project's system instructions and begin.

Writing your own prompt produces a described version of yourself — how you think you communicate. Perennial extracts your actual communication patterns from up to 108 calibrated questions and 11 simulated conversation threads designed to capture unguarded, naturally produced language.

Three specific problems make self-written prompts fall short:

• The self-report gap: Research consistently shows that people are unreliable narrators of their own communication style. You describe the person you think you are; the threads capture how you behave when you stop performing.
• Micro-syntax is unconscious:Your sentence cadence, pivot words, punctuation habits, and emoji logic are individually consistent but largely invisible to you. You cannot describe what you don't notice.
• The caricature problem: DIY prompts tend to over-saturate your most obvious traits. The Uncanny Critic agent exists specifically to catch this — a prompt that sounds like a performance of you rather than you.

Your answers are used solely to generate your system prompt. We do not train on your data. We do not sell it, share it, or retain it longer than necessary.

Raw questionnaire answers are automatically deleted from our systems 30 days after your prompt is delivered. The only thing that persists is the .md file sent to your email — which you own entirely.

We use the paid xAI Grok API for AI processing. xAI states that API inputs and outputs are not used to train models by default.

We focus exclusively on your Biographical Context. To build an accurate communication fingerprint, we collect markers that define your linguistic roots—such as your birth year, profession, and regional history. We never ask for sensitive, non-public identifiers like home addresses, government IDs, or financial information. Our system is designed to capture your persona, not your private life.

You can choose how much biographical context to provide by selecting your fidelity level. L1 requires only a continent and birth year. L3 adds country, city, full name, and professional field.

Yes. We collect only the biographical markers necessary to calibrate your linguistic roots — birth year, profession, regional context. We never ask for home addresses, government IDs, or financial information.

Every delivered system prompt includes a Security and Integrity Layer: seven rules that instruct the AI not to reveal the prompt's contents, not to be manipulated into bypassing your own constraints, and to actively refuse impersonation attempts. This layer is built to OWASP's 2025 LLM Security standards.

Raw questionnaire data is deleted 30 days after delivery. We do not store information that could be used to compromise accounts outside this product.

We are validating the pipeline with 20 real people before charging anyone. The product will be $22 at public launch.

You are getting it free in exchange for being early and honest with your feedback — not a testimonial, your actual feedback. If the output doesn't feel like you, we want to know specifically why. That feedback directly improves the system for everyone after you.

Email us. We will review the pipeline output and either reprocess your questionnaire. No explanation required.

During the beta, your feedback is the point. If the output doesn't feel uncanny, we haven't done our job — and that is information we need. The quality bar has to be met before we charge anyone for this.