Raw material accumulates in different forms. Sometimes it’s the weekly synthesis of reading and listening produced by my weekly review process. Sometimes it’s a transcript, a set of book annotations, or a cluster of notes that haven’t resolved into anything coherent. The material is there; what I’ve been missing is a structured way to make sense of it.
Recently I built a brainstorming command called /conclave. It runs inside Claude Code, a terminal-based AI interface. Point it at any source and it creates or opens a daily note in my writing vault, loads context from my recent work, and begins a conversation. Through a local server it has access to my full vault and Zotero library, giving it the ability to search notes, surface connections, and pull in sources directly.
The instruction behind the command tells Claude to push back rather than validate, to steelman positions before testing them, to ask one sharp question rather than a battery of gentle ones. It’s explicitly not a summarisation tool — building arguments with AI requires friction, not agreement.
What a session looks like
The format is simple. I write a few lines — half-formed, provisional — and Claude responds. I reply in bullet points. Claude asks something sharper. The session builds as a running transcript in the note.
The key mechanic is that Claude re-reads the full note each time it responds, rather than treating this as a linear back-and-forth. I can add to the note anywhere — at the end, or as bullet points slotted between earlier responses — and all of it becomes context for the next reply. The conversation branches: different threads pull in different directions before eventually converging on the core idea.
Here’s a short excerpt from a recent session. I opened with a loose thought about writing (the bullet points inbetween Claude’s paragraphs are my responses to that paragraph):
Michael: Blog post on getting joy from writing, then getting disillusioned, then realising that I enjoy expressing ideas, and I used to do that mainly through writing, and AI helps me do that better. Maybe not “joy”, maybe I mean something like “purpose”. I used to say that I feel compelled to write, not that I enjoy writing. Maybe that feeling of compulsion was coming from a need to articulate something I saw, and that I wanted others to see?
Claude:
The distinction between compulsion and enjoyment is doing a lot of work here — because they’re pointing at different things.
Compulsion usually means there’s something that won’t let you rest until it’s done. Enjoyment is about the process itself being pleasant. You’re saying the writing was never really the enjoyable part — it was the vehicle for something else: making a thing you’ve seen visible to someone else. The writing was just the best tool available for that.
- Yes, the enjoyment came from seeing an idea, clearly expressed
But here’s where I’d push back a little: there’s something in the struggle of writing that shapes the idea itself. Writing isn’t just transcription — it’s discovery. The act of trying to articulate something you’ve seen often reveals that you haven’t quite seen it yet. Does AI-assisted expression give you that same loop? Or does it short-circuit it — producing fluency before the idea is actually ready?
- I’ve learned to look past the fluency of the AI output
- In some important sense, I don’t trust the words
- I’m looking for the expression of the idea in the words
- There remains a productive and generative struggle
Later in the same session, I asked it to search my vault and Zotero for relevant material:
Michael: Explore my permanent notes and Zotero for some concepts to add weight to the piece.
Claude:
The most directly useful vault note is the Feynman quote in [writing is thinking]: “writing was not merely a record or trace of what he thought, but rather the medium through which he thought that enabled him to think more consciously.” That’s almost exactly the claim you’re making about AI — except the medium has shifted from inscription to dialogue.
From Zotero, Nguyen, Hong, Dang & Huang (2024, Studies in Higher Education) studied doctoral students working with AI and found that those who engaged in iterative, highly interactive back-and-forth with the tool outperformed those who used it as a supplementary information source. The empirical finding maps directly onto your practice: the collaboration is the thing, not the offloading.
[scarcity inversion when words become cheap] is also worth pulling in here. When generation becomes cheap, discernment becomes the scarce and valuable thing. Carlo Iacono: “when words get cheaper, meaning gets more expensive.” The vibe coder has outsourced the expensive part.
That last move — surfacing three things I’d written or read separately and hadn’t yet connected — is what makes this different from a conversation I’d have in a browser. The command has context because it knows what’s in my Obsidian vault and my Zotero library.
Once I feel like I’ve explored the idea sufficiently and moved the thread towards something resembling closure, Claude first generates new permanent notes based on the discussion, integrating them into my existing set of notes, and adds a conclusion to wrap things up. For example, this is the conclusion to the writing session described above:
Why this matters
Most AI interaction happens in a chat interface: you type, the model responds, you close the tab, the conversation disappears. If you want to add context, you have to search for it yourself (I do this as well, when I know what the connection is). If something useful emerged, you have to copy it somewhere manually.
In addition, I can swap out models in the background. I can switch between Claude Opus, Sonnet, and Haiku, or experiment with Codex, Gemini, or Qwen. Having a system that’s AI-agnostic means I’m not reliant on any provider and can switch to a completely different model at any point, without losing any context. It also means I can run a local, open-source model, although none are at a point right now where I’d do this permanently.
A session like this one is a markdown file on my machine from the moment it starts. It’s in my vault, alongside everything else, searchable, linkable, and available as context to any future session. When I asked the command to search my notes, it could find this session in a future search, surface what came out of it, and build on it. The thinking accumulates rather than evaporating.
That’s not a minor convenience. It’s a different relationship with the tool. The conversation isn’t a transaction — ask, receive, close. It’s a record of thinking, saved in the same place as the rest of my thinking, that compounds over time.