llm A field note by Vikrant Sharma
Thomas Ptacek's LLM writing workflow is the opposite of what I expected
The Matasano founder uses Claude to draft, then rewrites everything by hand. No copy-paste. The LLM is a research assistant that never ships.
Thomas Ptacek published his LLM writing workflow and it is not what I thought it would be. He uses Claude to generate first drafts of entire posts, then throws the draft away and rewrites the whole thing from scratch. No copy-paste. The model never ships. The workflow is brutal. He feeds Claude a topic, gets back fifteen hundred words of slop, reads it once, closes the window, then opens a blank document and writes the post himself. The LLM draft is scaffolding for his thinking, not a template to edit. This is the opposite of how most people use these tools. The default pattern is to generate something passable, tweak a few sentences, ship it. Ptacek treats the model like a junior researcher who hands you notes you will never quote directly. The value is in forcing yourself to articulate what you actually think, not in saving keystrokes. He also uses it for research. Ask Claude to explain a concept, get a rough outline, then go read the primary sources yourself. The model gets you to the right papers faster than Google, but you still read the papers. The generated explanation does not make it into the final post. The part that surprised me is he claims this is faster than writing from scratch. Staring at a blank page is slower than staring at a bad draft, even if you discard the draft. The LLM breaks the activation energy problem. You see what a mediocre version looks like, realise you can do better, then do better. I have been using models to edit, not to draft. This flips that. Generate the thing you will replace, not the thing you will ship. The model is a sparring partner, not a ghostwriter. I might try this for the next technical post I write and see if it actually saves time or just feels like it does.
Source: How to Write with an LLM
