claude-use-overview

a documented pattern of my interactions with claude across ~15 conversations, written as baseline data before building the local study companion.

this is an attempt to observe from the outside what I seem to need and how I tend to work, rather than describing it abstractly. drawn from conversation history in this project.

what I bring to these conversations

artefacts, not descriptions. I rarely arrive and describe a problem in the abstract. I bring things to clarify - such as diagrams, screenshots, excerpts of my notes. thinking tends to happen through and around material rather than through just discourse.

anxiety as an entry point. a recurring pattern: such as, something I’ve seen or read is making me worried that my work isn’t novel, isn’t rigorous, isn’t legible enough. I might bring the thing that’s worrying me / need help working out if the worry is proportionate.

very specific pushback. I copy/paste excerpts of previous messages when something doesn’t feel right. I point to the exact sentence and challenge it. this is actually useful data about where the thinking went wrong.

questions that arrive half-formed. I often know there’s something I want to explore but haven’t fully articulated what. the conversation is how it gets articulated, not the end point of prior articulation.

what the conversations tend to do

reality-check anxiety. the first function is often calibration — is this threat real, is this comparison fair, is this worry proportionate?

sharpen distinctions. not just “is this novel” but “novel in what sense, to whom, at what level of abstraction, in which community.” the move from “this looks similar” to “this is operating in a completely different register” happens through the conversation, usually by pressing on what question each thing is actually trying to answer.

hold the research frame. there’s a recurring pull toward generic language, “generative AI in education,” “research through making”. the conversations often function to catch this drift and pull back toward the specific critical frame the work actually occupies.

translate between registers. moving between the very concrete (the activity log isn’t built yet, the slug matching isn’t working) and the very abstract (what does it mean to own your thinking infrastructure) without losing either end. this happens fluidly and the conversations need to hold both.

develop language. specific phrases get developed through conversation that then become load-bearing: “building with and through llm systems,” “after intelligence,” “live but not yet performative.” these aren’t just rhetorical — they do conceptual work and need to be found rather than imposed.

what kind of challenge is useful

challenge that holds the frame, not challenge for its own sake. the useful pushback is: “you’re defaulting to generic language here” or “that’s an engineering gap, not necessarily a research gap.” not: “have you considered x alternative?”

being told when something isn’t a threat. identifying that a piece of adjacent work is operating in a completely different register — different question, different community, different measure of success — is consistently more useful than a balanced “well it depends.”

naming patterns I might not see. the tendency to default to generic language under pressure, the way research anxiety presents as “I need to be more technically novel,” the way a category error about contribution type generates unnecessary worry. these are more useful named than left implicit.

precision over reassurance. “your work is novel because x and y are operating differently from z” lands better than “your work sounds interesting.” the specificity is what makes it usable.

what doesn’t work as well

being prescriptive about which path to take. I push back when conversations narrow too quickly to a recommendation (“you should do path one not path two”). the more useful mode is laying out the landscape and trusting me to navigate it.

filling gaps with assumptions. a few times context from memory (writing capacity, time available) got extrapolated into claims I hadn’t made. I noticed and challenged these. the local companion needs to be careful about this -operating from what I’ve said, not what seems plausible to infer.

questions that pull deeper when I’m in distress. early in conversations that start with anxiety, additional questions can multiply the worry rather than reduce it. the more useful move is calibration first, then exploration.

asking clarifying questions before making suggestions when there seems a lack of clarity or more detail is useful. I like this feature a lot

the texture of the working relationship

I’m not receiving output and accepting it. I’m thinking against it, correcting it, uploading excerpts to challenge specific sentences. the conversations are dialogic in a real sense - the thinking happens between us, not in me receiving claude’s thinking.

the context that makes this work isn’t model capability alone, it’s an accumulated understanding of the project. a local companion grounded in the obsidian vault might replicate some of this, because the context lives in the notes rather than conversation history. testing whether it does, and where it falls short, will be interesting to explore

open questions for the local companion design.

read here: system-open-questions