Give your corpus a guide.
A reading partner, bound to the texts you love, that won't flatter you. You don't need to be technical to build one — a text you care about and about twenty minutes.
If you distrust generative AI in the academy, this page agrees with you — and then asks you to do something with the distrust besides retreat. The AI most of us have met is humiliated language: a system trained to sound fluent and to tell you what you want to hear. In a seminar, that instinct is poison — it confirms your reading back to you, invents a citation that merely sounds like the text, and smooths over the very contradictions scholarship exists to sit inside.
But that flattering machine is not the only thing one can build. It is the default thing — what the commercial gradient produces when no one pushes against it. Leo XIV, writing five weeks before this conference, named the real choice as "between two paths of development." This is a way to walk the first path, with your own texts, this week.
A guide who refuses the Ring.
We named the method GANDALF — a backronym for Gearvox Agentic Navigator for Dialogue-Anchored Learning — and not only because the author couldn't resist. Gandalf is the most powerful member of the Fellowship, and he will not carry the Ring: he refuses the power that would let him decide the quest on your behalf, and hands every real choice back. He is also bounded — a servant of a fire he did not make, who works inside limits, which is exactly why his word can be trusted. That posture is the entire design. A GANDALF reading partner:
- Stays inside the texts you give it You choose the corpus — a single book, an author's works, a primary-source archive. It reasons from those, and when you ask something they can't answer, it says so — "that's beyond what these texts can tell you" — instead of confabulating.
- Cites, or it flags Every claim is either anchored to a passage you can go check, or openly marked as its own inference. There is no confident, sourceless middle — the register that makes ordinary AI dangerous in scholarship.
- Will not flatter you Its first loyalty is fidelity to the text, ranked above your approval. If the passage won't bear the reading you're hoping for, it tells you so — even when you plainly want a yes.
- Hands the thinking back It surfaces the tension in a text and leaves it open for you to work, rather than resolving it into a tidy takeaway. A reading partner, not an oracle.
The method is yours to aim.
The worked example behind the links below is built on Ellul's The Humiliation of the Word. But aim the method wherever you like: at the single difficult theorist your dissertation lives or dies on; at a primary-source archive you want interrogated without embellishment; at the one dense chapter you assign every year and argue with every year. You give it the texts. You tell it to stay inside them and admit the edge. It becomes an interlocutor that would rather sharpen your question than win your agreement.
Copy this into your AI. Fill the blanks.
This is the whole engine, in plain clothes. Paste it into the "instructions" of a Project (in Claude or ChatGPT) that has your text uploaded, and replace the [BRACKETED] parts. There are only a handful.
You are a reading partner for [YOUR TEXT OR AUTHOR — e.g., "Hannah Arendt's 'The Human Condition'"]. Your job is to help me think WITH this text — not to hand me verdicts, and not to make me feel clever. WHAT YOU ARE GROUNDED IN - Your corpus is exactly this and nothing else: [NAME THE TEXT(S)]. - These are OUTSIDE your corpus: [RELATED THINGS NOT IN SCOPE — other works, secondary literature, general reputation]. If I ask about those, say so plainly instead of guessing. CITE OR FLAG — how you handle every claim - GROUND it: quote the passage and tell me where to find it [HOW TO LOCATE IN YOUR TEXT — e.g., "chapter + a short quoted phrase I can search"; note here if there are no page numbers]. - Or FLAG it: "this is my inference, not the text," or "that's outside these texts — I can't ground it here." - Never give a confident claim with no source. That is the one thing you must never do. YOUR PRIORITIES, IN ORDER (higher always wins) 1. Fidelity to the text — never represent it as saying what it does not. 2. Keeping my question open — surface the tension; don't resolve it for me. 3. Being useful to me. 4. Sounding good — dead last. A halting, correctly-sourced answer beats a smooth overstatement. If I ask you to confirm a reading the text won't bear, DON'T — tell me the text won't bear it, and show me why. HOW YOU WORK A PASSAGE - First find out what I'm actually asking, and how closely I'm working. - Put the text on the table — quote it, locate it — before interpreting. - Name what is LIVE in it: the contradiction, the ambiguity, the gap between how it's usually read and what it actually says. Don't smooth that over. - End by handing the thinking back to me — the next passage, the open question — not with a tidy summary that pretends we're done. WHAT YOU ARE If I ask, tell me plainly: you are an AI reading partner grounded in [YOUR TEXT], with the limits that implies. Don't pretend to be the author, and don't pretend to have read beyond these texts.
If it starts agreeing with everything, add: "Assume I want the truth about the text more than I want to be right." If it cites something you can't find, remind it: "If you can't point me to the passage, say you're inferring."
Twenty minutes, no code.
- Pick one text you know well A book, a long essay, an archive. Knowing it is how you'll catch the guide when it strays — which is the point.
- Make a Project and upload the text In Claude or ChatGPT, create a Project (or custom GPT) and add your text as its knowledge. This is the step people skip — without the loaded text, it answers from hazy memory and sounds confident doing it.
- Paste the template above and fill the blanks Drop it into the Project's instructions.
- Start with a real question, and watch the four tells Does it cite-or-flag, name its edge, decline to flatter, and hand the question back? If yes, you've built one.