Long-Form Writing Workflow

This is the general workflow I follow for articles, essays, and longer writing. It's still evolving -- I'm actively experimenting with which parts the agent handles well and which parts I need to do myself. But the overall shape has been pretty consistent.

1. Capture the Idea

It usually starts with a spark -- something I've been thinking about, a connection between things I've read, a take I want to flesh out. I either create a page for it directly in my LogSeq vault or send my agent a voice memo explaining the idea and have it create the page for me.

The voice memo approach works well when I'm not at my computer. I'll ramble for a few minutes about what I'm thinking, and the agent transcribes it and turns it into a structured page in the vault with the core idea, initial thoughts, and any angles I mentioned.

2. Research and Gather Context

Once the idea has a page, I pull in related material. I'll ask my agent to search the vault for notes related to the topic -- literature notes, reference notes, anything relevant from the zettelkasten. It links those to the article page so I have a research foundation to work from. A lot of this happens over voice memos -- I'll send a quick audio message like "search the vault for anything related to X and link it to the article page" and the agent handles it. (See Audio Transcription for how to set up voice input.)

Sometimes this surfaces connections I hadn't thought of, which is the whole point of having a linked knowledge base. A note from one book connects to an insight from a completely different paper, and that becomes part of the argument.

3. Outline

This is where the back-and-forth with the agent is most useful. A few ways this goes:

• Interview style -- I ask the agent to ask me questions about the idea. I respond via voice memos, and it updates the page with my answers organized into a structure. This is great for pulling ideas out of my head that I haven't fully articulated yet. Most of my outlining happens this way -- talking through ideas is faster than typing them out. (See Audio Transcription for setting up voice input.)
• Solo in LogSeq -- Sometimes I just sit with the page in LogSeq and write out ideas, move things around, build the structure myself. The agent isn't involved in this part.
• Collaborative -- I write a rough outline, send it to the agent, and we go back and forth refining it. It might suggest reordering sections or point out gaps in the argument.

The goal is a solid outline before any real drafting starts. The outline is where the thinking happens -- the draft is just filling it in.

4. Drafting

This is the part I'm still experimenting with. Two approaches:

Agent-generated first draft -- I give the agent the outline and ask it to write a full draft using my voice profile. This works okay. The structure is usually right and the ideas are there, but the voice still needs work. Even with a strong VOICE.md file, the output reads more like "well-prompted AI" than like something I actually wrote. It's a decent starting point though -- better than staring at a blank page.

Writing it myself -- Sometimes I just write the first draft from the outline. This produces better writing but takes more time. I'm honestly still going back and forth on which approach I prefer. Ideally the agent would generate a draft good enough that I only need to edit, not rewrite. We're not there yet, but it's getting closer.

5. Editing

This is mostly me. I've tried using the agent for editing and haven't found a way to make it consistently useful. It's good at catching structural issues (sections that don't flow, arguments that need support) but not great at the sentence-level stuff that makes writing sound like a specific person.

So I do the editing pass myself -- tightening sentences, cutting filler, making sure it sounds like me. This is the part that takes a piece from "AI-assisted" to "actually mine."

6. Publishing

Still manual. I copy the final version and publish it myself, whether that's to a blog, Substack, or wherever. I'd like to automate this eventually -- have the agent handle formatting and publishing to different platforms -- but I haven't built that yet. It's on the list.

What Works Well, What Doesn't

Works well:

• Capturing ideas via voice memo
• Research and gathering related vault notes
• The interview-style outlining process
• Generating a structural first draft from a good outline

Still figuring out:

• Getting the agent to produce drafts that genuinely sound like me (see Voice Profile for more on this)
• Using the agent for editing
• Automated publishing

The biggest lesson: the agent is most useful in the early stages (capture, research, outline) and least useful in the final stages (editing, voice). That might change as the models improve, but for now the workflow reflects that.