Ruben Hassid is an AI writer with a large readership on X, with more than 500,000 people reading his newsletter every week. In a long tweet thread, he shared a methodology: with two hours and a single text file, you can get any large language model to write exactly like you. This article will break down the complete workflow of the methodology and evaluate its real-world value and limitations.

The Core Proposition: Why a Text File Can Replicate Your Personality

Hassid’s argument is direct and provocative: you are nothing more than a text file. He believes most people overestimate their complexity and underestimate the extractable patterns in their writing. Your voice, your taste, the terrible posts that make you cringe, the phrase an old friend would use when imitating you, the two characters you typed and deleted, the metaphor you’ve used three times this year without realizing it — every one of these is a capturable pattern.

The underlying logic of this view rests on the in-context learning capabilities of current large language models. When a model reads a carefully crafted personal profile file at the start of a conversation, it keeps referring to those constraints throughout the generation. This isn’t training a model; it’s handing the model a “character script” that makes it play you in every output.

Hassid’s own validation is quite persuasive. He has 20 years of writing experience, has lived in Seoul, Berlin, and Tel Aviv, learned English at age 9 from a gaming forum, dropped out of college twice, and now does AI consulting for Fortune 500 companies. After compressing all that experience into a file, Claude’s first drafts came out “as if I’d written them myself,” sometimes even “written before I’d thought of it.”

The Two-Phase Structure of the Methodology

Hassid’s methodology is divided into two tightly connected phases: the deep interview and the compression compilation.

Phase One: A 100-Question Deep Self-Interview

The core tool of this phase is a Claude prompt Hassid calls the “Taste Interviewer.” Its task is to systematically extract your writing DNA through 100 questions. The questions are divided into seven major categories:

The interview rules are strictly designed: ask only one question at a time; push back on vague answers; demand concrete examples; point out contradictions; dig deeper into interesting topics. Hassid recommends using Wispr Flow for voice input, because “speech is faster and more honest.”

Completing all 100 questions takes about 90 to 120 minutes, and the final output is a raw profile of around 20,000 characters. The full interview prompt is in Appendix A at the end of this article; you can copy it directly into Claude.

Phase Two: Compressing Into a High-Fidelity Profile File

The problem with the raw profile is its size. Having Claude read 20,000 characters in every conversation consumes context window and increases token costs. The solution is to have Claude act as a “Voice Compiler” that compresses the raw profile into an about-me.md file of 2,000 to 4,000 tokens.

The core rule of compression: every line must pass the test — “If this line disappeared, would the AI’s writing, editing, judgment, refusals, structure, or decision-making be different?” If the answer is yes, keep it; otherwise delete it.

The compressed file uses an XML structure containing 15 modules. Each module has a clearly defined functional boundary:

• usage: three lines of instructions telling the AI how to use the file
• priority: six priority rules, the decision logic for handling conflicts
• identity_context: identity details that influence voice, taste, metaphors, and judgment
• voice_fingerprint: actionable descriptions of rhythm, density, directness, humor, emotional temperature, etc.
• writing_laws: specific writing rules, including positive and negative examples
• communication_laws: rules for emails, replies, requests, criticism, and other scenarios
• hard_refusals: things the AI should never write, say, or imply
• taste_loves and taste_disgusts: specific things you appreciate and loathe
• phrase_bank: words, phrases, and metaphors you commonly use, plus expressions you avoid
• signature_tells: small details that make you recognizable
• decision_rules: how you judge quality, honesty, beauty, risk, and credibility
• productive_contradictions: tensions to preserve rather than smooth over
• golden_examples: 3 to 6 contrasting positive/negative examples
• do_not_infer: assumptions the AI should not infer from this profile
• final_instruction: a concise instruction telling the AI to silently apply this profile

The full compression compilation prompt is in Appendix B at the end of this article.

Real-World Results: From Experiment to Workflow

Hassid demonstrates several comparison cases. Before and after adding the about-me file, the same AI model’s output quality on the same prompt differs significantly. Without the profile, the model outputs generic, safe, characterless text; with the profile, the output carries clear personal style, specific word choices, and structural preferences.

The more practical value lies in team scenarios. Hassid points out that you can send this file to team members so they can draft content in your style even when you’re not around. Customer-service replies, email communication, and content creation can all be handled by the team while maintaining a consistent voice.

However, Hassid also candidly acknowledges one problem: consistency brings predictability. Once your style is fixed in a file, your output patterns become predictable too. His solution is to edit the file regularly, because a person’s taste and voice change over time.

The Value and Limitations of the Methodology

Why This Method Works

The cleverness of the methodology lies in breaking the vague goal of “making AI write like me” into two executable concrete steps: first, a structured interview forces you to make your writing rules explicit; second, compression filters out the noise and keeps the signal.

Most people can’t write a good about-me file directly because they’ve never systematically thought about their writing preferences. The 100-question interview process is essentially forced introspection — it pushes you to convert taste at the intuitive level into articulable rules. This conversion is precisely the prerequisite for AI to understand and execute.

Limitations and Things to Watch Out For

First, the method assumes your writing style is stable and patternable. For writers whose style is highly variable and experimental, this profile could become a constraint instead.

Second, the return on the 2-hour investment depends on how often you use it. If you only occasionally ask AI to write a few emails for you, spending two hours building a profile may not be worth it. But for people who need continuous content output — newsletter authors, social media operators, team managers — this one-time investment can bring long-term efficiency gains.

Third, Hassid’s method relies heavily on Claude’s Opus model and Extended thinking mode. On other models or cheaper models, interview depth and compression quality may drop noticeably.

Finally, and Hassid himself stresses this: the process can make people uncomfortable. When you read yourself into a text file, there’s “nowhere to hide.” Every belief is a commitment; every refusal is a rule you must now follow. He “recoiled” the first time he read his own compressed file.

How to Get Started: A Minimum Viable Practice

If you don’t want to go through the full 100-question process at once, you can start with a minimal version:

1. Open any AI conversation and paste in Hassid’s Taste Interviewer prompt (see Appendix A)
2. Answer only the first 20 questions, focusing on the “writing mechanics” and “aesthetic crimes” categories
3. Have the AI compile the answers into a lean profile of 500 to 1,000 tokens
4. The next time you need AI to draft content, put this profile at the start of the conversation
5. Gradually supplement and correct the profile based on output quality

This minimal version takes about 30 minutes and is enough to experience the core value of the methodology: the qualitative leap in AI output from “generic” to “like you.”

References

• Ruben Hassid’s original tweet thread
• Claude official documentation
• Wispr Flow voice input tool

Appendix A: Complete Taste Interviewer Prompt

Paste the following into a Claude conversation to start the interview:

Appendix B: Complete Voice Compiler Prompt

After completing the 100-question interview, paste the following into the same conversation to have Claude compress the raw profile into a high-fidelity about-me file: