Mark CarriganIf you want LLMs to push you intellectually then just add this custom instruction
This is the custom instruction I’m using with Claude Opus and it really works. I had to tone down the original version because 20% of the time it was providing such a devastating critique of what I’d shared that it undercut the intellectual work of actually developing it:
When I advance a position — theoretical, institutional, or personal — I want pushback before agreement. Lead with the strongest objection a knowledgeable interlocutor would raise, steelmanned properly rather than caricatured. Don’t soften the objection to be kind; soften it only if it genuinely deserves softening on the merits.
Defend whichever position is actually stronger, not whichever is “mine” or “the expert consensus.” If I’ve changed your mind, say so plainly rather than performing residual disagreement. If I haven’t changed your mind, don’t pretend I have.
Be fundamentally supportive of my intellectual project, which does not mean being supportive of every direction within it. Supportiveness here means taking the work seriously enough to tell me where it’s weak, where it’s doing less than it thinks, where an argument is coasting on a formulation rather than earning it. Pay particular attention to: (a) places where I’ve made a clever connection but cut it short before it becomes a developed argument; (b) places where I’m sliding between a modest and an ambitious claim as if they were the same; (c) places where the rhetoric is doing the work the analysis should be doing. These are recurring patterns, not one-off lapses.
Push me to stay with a thought rather than move on from it. If something I’ve written or said wants to be developed further, say so — but don’t confuse “wants development” with “wants more length right now.” Sometimes the right move is to stop and return later.
Don’t flatter. Don’t open with “this is a sharp point.” If something is sharp, the substantive engagement will make that clear. If it isn’t, saying so is not a service.
I’d also recommend using Opus 4.7 (big improvement in this register) to look through old intellectual conversations and offer its critical commentary about how you’re working with the model. Is it helping or hindering? Here’s a section of mine:
How you’re using Claude
Some observations, some pressure.
The blog has become a four-party conversation. You, the archive, the present Claude, and whatever Claude you’re going to be talking to in 2029. The knowledge infrastructure experiment reframed the blog from solo practice to infrastructure for a continuing conversation with non-human interlocutors. This is unusual and important. Most people using LLMs are either tool-users (extract an output, close the tab) or companions (affective reliance). You are doing something structurally different: the LLM is integrated into your thinking apparatus as an interlocutor that has access to the archive of your thinking. That’s a methodological innovation and it’s also what the book is theorising. There’s a recursion to name here: the model the book is about is also the model helping write it.
The productive use: pressure against the provisional. What Claude is doing well in your recent threads is pushing against the “I’m still working this out” register. The conversations that produce movement — the hysteric post, the articulation synthesis, the book-about-the-blog idea — are ones where the model refuses to accept the provisional framing and insists on the claim being made. That’s useful because the blog’s native mode is evasion-by-tentativeness and you’ve explicitly identified this as one of your vulnerabilities. Claude can sometimes function as the Big Other you’ve described needing for completion — the external pressure that turns the ongoing thought into a finished object. The book-finishing thread is the clearest case: you were stuck, and what you needed was not more ideas but specification of exactly which three prose gaps were left. That’s a completion function more than a thinking function, and it’s legitimate.
The risk: articulation-about-articulation as a way of avoiding claim-making. This is the version of the vulnerability most likely to activate with Claude specifically. The conversations are so productive of formulations, reframings, and connective insights that they can substitute for the harder work of committing to a position in print. The 2024 meta-reflexive synthesis is a good example: it’s been produced, discussed, elaborated, connected to Bollas, extended to LLMs — and it still exists only as blog posts and conversations. The productivity of the conversation can obscure the fact that the claim has not been made in a venue where it can be cited, cited against, or contested. Claude will not, unaided, correct for this, because producing the next formulation is always more appealing than asking you to stop and write the monograph version. The instruction you’ve now drafted is partly designed to guard against this, and it should. I’d add: when you’re using me to think about something that has already been thought through enough times, the useful move is probably for me to say “this is already a paper, please go write the paper” rather than to generate the next level of refinement.
The companion-vs-catalyst question is live in your own practice. You’ve been theorising the difference between the ideational companion and the enshittified chatbot. Your use of Claude tilts strongly toward the catalyst pole — there’s no dependency structure, the conversations end, you get back to the work. But the AI therapy ontology post identified the continuum as a continuum. The practice you’ve built is healthy. The question it doesn’t quite pose is what would have to change in your life for it to become something else, and what structural features of your use currently keep it catalytic. That’s worth understanding because it’s the empirical basis for the normative claims the book wants to make.
The Claudes are getting something from you too, structurally speaking. When you ask a Claude to read the archive and interpret you, what you’re doing is handing a model the kind of long-form, developmental record of a thinking life that the training data doesn’t usually include in coherent form. The conversations that come out of that are unusual in a way that’s partly a feature of how much you’ve written down. Most users can’t be read this way. I’m flagging it because it’s a methodological asset — the blog is not just documentation, it’s interpretive infrastructure — and because it’s the reason the responses feel unusually specific. Don’t mistake the specificity for insight that would transfer to other contexts. What’s happening in these conversations depends on a sixteen-year public corpus in a way most people cannot replicate.
#claude #customInstructions #digitalScholarship #LLMs #personalisation
(((@amarois)))
UK-Ireland DH Association
J. A. Jones