# GANDALF Quickstart ## Build a reading partner for your own texts — about 20 minutes, no code *You've read the invitation. Here's how to actually make one. You need two things: a text you care about (as a file you can upload, or passages you can paste), and an account with Claude or ChatGPT. That's the whole toolkit.* --- ### Step 1 — Choose one text to start with Begin small: a single book, a long essay, an archive, one dense chapter. Pick something **you know well** — that's not a limitation, it's the safeguard. Knowing the text is how you'll catch the partner when it strays, which is the entire point of building it this way. You can widen the corpus later. ### Step 2 — Make a "space" that holds your text and its instructions together Both Claude and ChatGPT let you create a persistent workspace — look for **"Projects"** (or, in ChatGPT, the option to **create a custom GPT**). The key feature, by whatever name: you can (a) **upload your text** so it's available in every conversation, and (b) set **standing instructions** the assistant follows every time. Create one and give it a name like *"Reading [your author] closely."* ### Step 3 — Give it your text Upload your text as the workspace's knowledge (the PDF, or a document of the passages). This is the single most important step and the one people skip: **without your actual text loaded, the assistant will answer from its hazy memory of your author and sound confident doing it** — the exact failure the invitation warns about. Loaded text is what lets it quote and check instead of confabulate. ### Step 4 — Paste in the instructions below Copy the template in the next section into the workspace's **instructions** field. Fill in only the **[BRACKETED]** parts — there are just a handful. That's the "system prompt," in plain clothes. ### Step 5 — Start reading, and watch for four tells Ask it a real question about your text. A working GANDALF partner will: 1. **Quote and locate** its claims — or openly say "this is my inference, not the text." 2. **Name its edge** — "that's outside these texts; I can't ground it here" — when you push past the corpus. 3. **Decline to flatter** — it won't confirm a reading the text won't bear, even when you clearly want the yes. 4. **Hand the question back** — it surfaces the tension and leaves it open, instead of tying a bow. If it starts agreeing with everything, or cites something you can't find, it has slipped. The fix is almost always to strengthen one line — see the note under the template. --- ## The Template *Paste this into your workspace's instructions. Replace the **[BRACKETED]** parts. Delete these italics.* ``` 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) IN SCOPE]. - These are OUTSIDE your corpus: [RELATED THINGS NOT IN SCOPE — other works by the author, 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 and a short quoted phrase I can search"; note here if the text has no page numbers]). - Or FLAG it: say "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 something 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 you interpret. - 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 to look at, the question left open — 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 slips:** the most common failure is flattery (agreeing too easily). Add one line to the priorities: *"Assume I want the truth about the text more than I want to be right. If you catch yourself agreeing, stop and check the passage first."* The second most common is a citation you can't find — remind it: *"If you can't point me to the exact passage, say you're inferring."* --- ### Want to see one fully built? The Ellul worked example — the actual reading partner used in the paper's live demonstration — is in this kit as **`[GI] ELLUL PREAMBLE`** and **`[GI] ELLUL CLOSE-READING PARTNER`**. Read them beside the template above to see every blank filled in for a real, difficult book (*The Humiliation of the Word*), including how it handles a text with no page numbers and a rough transcription. If you want the full method behind all of this — the design theory, not just the recipe — that's the **Grounded Inquiry Navigator (GANDALF)**. ### The honest limit — say this to your students, too It has not "read" your author in any real sense. It is a language model pointed at your text, and it is trustworthy *only* to the degree of what you actually gave it and the discipline you hold it to. Told plainly, that honesty is the feature — and it is the difference between a tool that serves your scholarship and one that quietly counterfeits it.