TL;DR
The thinking of anyone who writes a lot and systematically can be extracted from their texts and encoded into a prompt — so the LLM starts reasoning with their logic, not just imitating their style. The researchers call this the Relic Condition: a mechanism by which a public body of writing becomes raw material for building an "intellectual double."
Key discovery: every serious author has a stable architecture of judgment — how they define concepts, which distinctions they treat as decisive, which arguments they reject as category errors, what they lean on. That architecture is invisible in conversation but clearly visible in their writing: publication pressure and peer review literally force authors to make their logic transparent again and again.
The method works in two steps: first you extract the expert's reasoning system from the corpus — their analytical moves, evaluation thresholds, favorite distinctions, refusal patterns. Then you encode that into structured constraints for the LLM and get an assistant that answers not like "smart ChatGPT" but like a specific thinker applied to a new problem.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
STEP 1: Collect the corpus — the expert's articles, books, chapters
→ Load into the LLM (one chat or via file upload)
STEP 2: Extract the reasoning system — ask the LLM to identify:
- How the expert defines key concepts
- Which distinctions they consider critical
- Which arguments they reject, and why
- Which theories and ideas they rely on by default
- How their logic of proof is structured
→ Get a structured "thinking portrait"
STEP 3: Build the constraint prompt — lock in the architecture
of thinking as a set of explicit instructions for the LLM
→ Get an "expert bot"
STEP 4: Pose a new problem — apply the expert bot to your question
→ Get an answer in that specific thinker's logic
All steps can be done in a single chat with file uploads.
Example
Task: You're building an edtech startup and want to take your idea apart the way Paul Graham would — the Y Combinator co-founder who has written hundreds of essays about startups, growth, and founder mistakes.
Prompt (Step 2 → 3):
I'll upload several of Paul Graham's essays and posts (text below).
Your task is to extract his reasoning system as a startup analyst.
Find and structure:
1. How he defines a "working idea" vs a "non-working" one — his criteria and thresholds
2. Which distinctions he considers decisive (e.g., pain vs want,
customer vs user, etc.)
3. Which arguments he automatically dismisses as weak —
his "refusal patterns"
4. Which concepts he relies on by default
5. How his logic for breaking down an idea is structured — where he starts,
where he goes next
After the extraction, become an analyst with this reasoning system
and take apart my idea: [describe the idea].
[Paste Paul Graham's texts]
Result: The model first shows a structured "thinking portrait" — 5-7 blocks with specific criteria and logic. Then it applies that logic to your idea: it points out weak spots using exactly the words and arguments characteristic of Paul Graham. The breakdown won't be generic "ChatGPT advice" but an analysis in a specific analytical tradition.
Why It Works
Why a plain "answer like Paul Graham" prompt doesn't work. The LLM knows the name and the general style, but not the architecture of judgment: how exactly this person defines concepts before the analysis begins, which arguments he considers category errors, what logic of proof he uses. Without that you get surface imitation, not thinking.
What LLMs do well. Models are excellent at recognizing patterns in large volumes of text. If one person has written 30+ long essays, their analytical moves repeat: the same distinctions, the same rejections, the same sequence of reasoning. An LLM picks out those patterns better than most readers.
How the method closes the gap. Instead of the command "be like X," you give X's explicit structure of thinking — instructions extracted from the texts themselves. The LLM stops guessing at style and starts following locked-in logic. That's exactly why, in the study, experts rated the result above "a strong academic in general" — they recognized a specific analytical tradition.
Control knobs: - Corpus size → the more texts, the more precise the extraction. Minimum: 5-7 long essays - Task specificity → the method works better on structured tasks (evaluate an idea, give feedback) and worse on improvisational ones (live dialogue, unformatted questions) - Explicit extraction → you can ask the LLM to show the thinking portrait before applying it — then you see it yourself and can correct it - Corpus type → works better with analytical texts (essays, breakdowns), worse with fiction or purely rhetorical writing
Prompt Template
Below are texts by {expert name}. Study them and extract their reasoning system.
**What to find:**
<ObjectDefinitions>
How {name} defines the key concepts in their field.
Find formulations that recur — these are their working definitions.
</ObjectDefinitions>
<CriticalDistinctions>
Which distinctions they consider decisive and return to again and again.
What do they separate where others lump things together?
</CriticalDistinctions>
<EvaluativeThresholds>
By what criteria they judge ideas, arguments, and decisions as strong or weak.
What quality thresholds do they use?
</EvaluativeThresholds>
<RefusalPatterns>
Which arguments they systematically reject.
What counts for them as a "category error" or an obviously weak move?
</RefusalPatterns>
<InferenceSequence>
How their line of reasoning is structured: where they start the analysis,
which questions they ask first, where they go next.
</InferenceSequence>
Once you've described the system, become an analyst with this
reasoning system and apply it to the task: {task}
---
TEXTS:
{paste the expert's articles, posts, chapters}
Placeholders:
- {expert name} — the person whose thinking you're cloning (Paul Graham, Seth Godin, Ben Thompson, any author in your field)
- {task} — what needs analyzing: a product idea, a strategy, a text, an argument
- {paste texts} — at least 5-7 long-form texts by the expert
🚀 Quick start — paste into the chat:
Here is a template for cloning an expert's reasoning system.
Adapt it to my task: [your task].
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask which expert to clone, which texts are available, and what needs analyzing — because it needs a corpus and a concrete task for the extraction to make sense.
Limitations
⚠️ Weaker at improvisation: structured tasks (evaluate a piece of work, write a review, take apart an idea) are the strong zone. Live dialogue and answers to unexpected questions are weaker.
⚠️ You need a big enough corpus: with 2-3 short texts, the extraction is superficial. The study used 35-68 analytical units. In practice — at least 5-7 long, varied texts by one author.
⚠️ It clones relational tacit knowledge, not everything: the method captures explicit reasoning logic. "Embodied" expertise (intuition, accumulated experience, a feel for the field) is outside its reach. But logic is exactly what most work tasks need.
⚠️ Narrow study sample: tested on two humanities scholars. Whether it works the same for entrepreneurs, engineers, or journalists needs separate verification.
Adaptations and Extrapolations
🔧 Technique: add an explicit verification step → check extraction accuracy
After the LLM extracts the reasoning system, add: "Now give 2-3 quotes from the texts that support each of the points." This makes the extraction transparent and lets you correct inaccuracies before applying it to the task.
🔧 Extrapolation: cloning an organization instead of a person
The same principle works for organizations with a consistent public stance. Upload a company's annual reports, public talks, and strategy documents — and extract the organization's reasoning system. Uses: modeling how a competitor will react to your move, how an investor will evaluate a proposal.
🔧 Extrapolation: cloning your own thinking to scale yourself
Gather your best writing, breakdowns, and decisions from the past year. Run the extraction on yourself. You'll get a structured description of your own analytical logic — and a prompt that lets the LLM answer within your frame of reference. Valuable for delegating tasks to an LLM without losing your style and approach.
How They Tested It
The team took two real humanities scholars with international publications — for the first they processed 68 analytical units (~1742 pages), for the second 35 sources (articles, books, chapters). From these corpora they used an eight-step method to extract each scholar's reasoning system and encoded it as structured constraints for an LLM with a high-reasoning config — in effect, they wrote prompts that define the architecture of a specific scholar's thinking.
The design is interesting: for testing they used functions where the standards are known — peer-reviewing papers, academic supervision, lectures, scholarly debate. Three independent senior academic experts wrote 18 evaluation reports and 6 syntheses in the format of "recommendations for appointment." In parallel, a test round ran the same tasks through plain ChatGPT/Claude/Gemini as a baseline — to check that the result wasn't just "a strong frontier chatbot."
The sharpest point: not one of the six independent evaluators rated any bot below Senior Lecturer level (roughly equivalent to a tenured Associate Professor in the US). The lowest rating came from the strictest evaluator — and his argument wasn't "the quality is poor" but "no publication record and grant history." In other words, the work was credited on substance; the cap came from formal credentials the bot simply doesn't have. That's telling: the bots didn't just "clear the bar" — they cleared it with room to spare.
What's surprising: different bots kept different personalities. The first was strong on conceptual boundaries and political analysis. The second on operationalization and pedagogy. No expert described them as equally "just a strong academic" — they recognized specific analytical traditions. That suggests the method captures an individual architecture, not averaged-out competence.
Resources
Paper title: The Relic Condition: When Published Scholarship Becomes Material for Its Own Replacement
Authors: Lin Deng (University of New South Wales), Chang-bo Liu (Independent Researcher)
Contact: [email protected]
Key references from the study: Polanyi (1966) — tacit knowledge; Collins (2010) — three types of tacit knowledge; Autor (2015) — complementarity of labor and automation; Crawford (2021) — AI and the extraction of human labor; CD Projekt Red (2020) / Gibson (1984) — the Relic concept as portable consciousness