Youāre Not Running Out of Claude. Youāre Just Burning It Wrong.
This is today's substack post. Still tryign to figure out how to link everything up: A writerās guide to getting more out of every session I used to think Claude was broken. Somewhere around message twenty-five, it would start getting weird. Shorter answers. Stranger suggestions. Iād be mid-scene, building real momentum, and suddenly my AI writing partner had the attention span of a golden retriever at a squirrel convention. A very expensive, very confused golden retriever. I blamed the tool. Classic mistake ā right up there with blaming the pan when you burn the toast. Then I actually looked at what was happening under the hood. Every message you send, Claude re-reads the entire conversation from the top. Message one costs one unit of processing. Message ten costs fifty-five. Message thirty? Nine hundred and sixty-one. The cost doesnāt climb in a straight line ā it compounds like a credit card you forgot existed. By the time youāre deep in a session, ninety-eight percent of your token budget is spent re-reading old context. Your actual question gets one and a half percent of the available brain. No wonder it starts writing your villain like a slightly confused golden retriever. Hereās the thing nobody tells you when you first start using AI for your writing: long conversations are expensive. Not just in the āyouāll hit a limitā sense, but in the āyouāre getting progressively worse outputā sense. Youāre not unlocking some hidden depth by going deeper into a single chat. Youāre usually making it worse ā like stirring a bowl of soup until itās just warm beige sadness. So. What do you actually do about it? 1. Start fresh more often than feels right. Writers hate this because it feels like losing progress. Youāre not losing anything. Copy Claudeās last output ā the scene it drafted, the outline it built, the character notes ā and paste it as the opening context of a new chat. āHereās where we left off. Now letās do the next scene.ā You carry the work forward without dragging the entire conversation history behind you like a suitcase full of every draft youāve ever written. 2. Front-load everything that matters. Before you write a single creative prompt, drop in your story context: genre, tone, protagonist, where you are in the story, what you need this session to accomplish. One block, upfront. Then ask your question. Claude doesnāt need to be reminded ten messages in who your main character is if you told it clearly at the start. Itās not goldfish-brained. Itās just expensive. 3. Be specific about what you want back. āWrite the next sceneā is an open invitation for Claude to go wherever its circuits feel like going. āWrite a 600-word scene where Mara confronts her sister in the kitchen, present tense, third person limited, ending on unresolved tensionā gives it a target. Specific prompts get specific results. One attempt instead of three ā which is also how you get better scenes and preserve your sanity simultaneously. 4. Phase your work across sessions. Donāt try to do research, outlining, character development, and drafting in one marathon chat. Thatās how you end up with a conversation eating its own tail while you stare at your screen questioning your life choices. Research session. Outline session. Draft session. Each one short, purposeful, and cheap. The handoff is just copying the output from one phase into the opening of the next. 5. Save the heavy lifting for when youāre sharp. Your Claude budget and your own brain work exactly the same way. Complex creative work ā building a plot structure, working through a characterās arc, drafting a difficult scene ā costs more than simple tasks like reformatting your notes or generating a chapter summary. Do the hard stuff first. Push the light maintenance work to later in the day when youāre running on fumes and your best creative decision is choosing which snack to eat.