Possible Paths

Harrowings might profit from some outside help

I’ve been turning this over since we looked at that note on Claude Code and the Model Context Protocol. The practical wisdom there about doing the foundational setup once so every later session goes more smoothly feels directly useful for the kind of writing I do on the Substack. Not as a replacement for the work, but as a way to remove unnecessary friction so the real inquiry and voice can move more freely.

The heart of it is treating the writing environment itself with the same care one might bring to any long practice. Instead of jumping into a new draft or revision while the model has only a vague sense of who I am on the page, I connect a few reliable bridges first. The filesystem connection lets the collaborator actually see the drafts, notes, research folders, and images where the work lives. The Substack-specific bridge, the one built for authenticated access to articles and notes, lets it pull published pieces, search the archive, retrieve full content, and even surface engagement patterns or comments when that serves the thinking. Those two connections alone change the texture of the exchange. The model stops guessing at context and starts working from the actual body of work already in motion.

Alongside the connections sits a persistent set of guidelines that travel with every session. I keep them in a simple document the model reads at the start of each interaction. They describe the voice I’m after, the recurring themes that matter, the standards for research and accuracy, and the boundaries around sensitive material. They also spell out the expected rhythm of a session itself. Before any substantial drafting or editing begins, the model is asked to read relevant existing pieces or notes and report back what it notices about voice, motifs, and how the current task sits inside the larger body of work. Only after that reconnaissance does the real collaboration open. It is a small discipline, but it prevents the drift that comes from working in the dark.

The effect on the writing shows up in several ways. Drafts arrive already closer to the register I actually use, with less need to strip away generic phrasing or restore particular rhythms. When I’m exploring a thread that runs through older posts, the archive is no longer something I have to manually reassemble in my head. Thematic connections surface more readily, which supports the longer intention of the Substack as a coherent body rather than a sequence of separate pieces. Research suggestions stay grounded because they can be checked against material already in the folders. And the whole process feels more like an extension of the same “good ancestor” orientation that runs through the writing itself. The scaffolding is built so the work can compound without constant re-explanation or loss of fidelity.

What used to feel like occasional helpful sparks becomes something steadier. Sessions begin with less waste and more shared ground. The model knows the difference between a quick note and a deeper excavation, and it respects the places where restraint or precision matters more than fluency. None of this replaces the slow, embodied work of sitting with the material. It simply removes some of the mechanical drag that used to interrupt the flow between one piece and the next.

I’ve been surprised by how much the initial investment changes the felt quality of the collaboration. The same principle that appears in recovery work or in any sustained practice shows up here: most of the roughness comes from skipped foundations. Once the connections are live and the persistent guidelines are in place, the later sessions carry more of the actual texture of the inquiry rather than fighting to recover it. It is infrastructure for the long game, not a shortcut around it.

If the shape of this approach feels worth trying, the pieces are straightforward to put in place. The filesystem and Substack bridges are the natural starting pair. The guidelines document can be written in one sitting and refined over time as the work itself clarifies what needs to be held steady. And the simple habit of asking for reconnaissance before generation turns out to be the hinge that keeps everything aligned.

It is one more way of bringing the same attention to the tools that I try to bring to the writing itself. The result is not louder or faster output, but cleaner transmission of whatever is actually moving through the work at any given moment. That feels worth the modest setup time for the assistant.