Same Model, Different Week: Building a Harness for Claude Over the Weekend

Zephyr breaks building a working environment for Claude into seven parts: four on Saturday, three on Sunday. A harness can be thought of as horse tack, fitted around the model so it can work inside your files, tools, and permissions. The goal is concrete: the next time you open a chat window, you don't have to start with a self-introduction.

Saturday tackles 'knowing.' CLAUDE.md holds who you are, how you work, and things it has never done. Skills are organized into folders by repeated work, recording the steps for each operation. The memory file stores decisions, names, and facts that need to be remembered across sessions. All three point to the same thing: take context out of chat history and put it into reusable files.

Sunday tackles 'touching.' Connectors connect email, calendars, drives, and actual apps. Permissions gate sending, deleting, paying, and publishing, and ask first. A checklist performs a second check before a draft reaches your hands. Schedules let tasks run on their own, reducing how often you have to 'open a conversation and then start it.' The last line in the image says it plainly: seven parts, and the model is the one everyone has.

Building it isn't the end. Zephyr's next step is to run it on real work and use mistakes to improve the tools, rather than continuing to add settings. If a draft doesn't feel right, add the phrase to the never list in CLAUDE.md. If you ask the same question twice, put the answer into the memory file. If the same type of task fails the same way, write the steps into a skill instead of leaving them in the chat. If it does something you didn't want, require it to ask first, and make that the rule every time. These four cases are more concrete than any 'best practice,' because they come from your own work record.

The image also includes a troubleshooting table. The 'knowing' stuff lives in CLAUDE.md, skills, and memory files. The 'touching' stuff lives in connectors, with permissions layered on top. Checking relies on reading the checklist you ran before. Running independently relies on scheduled tasks. If something breaks, first know which file to open. The comparison table is more direct: without this harness, every chat starts from scratch, copying and pasting in and out, you catch every error, and nothing runs unless you start it. With a harness, it starts from your files, works in real tools, asks before leaving, and checks itself first. The line at the bottom is: Same model. Different week.

Owen said in a reply that tutorial hell is real and you have to just do it. Kekko D'Amato called the seven-step framework the key, reasoning that fluency comes from repeated exposure to concrete workflows; abstract knowledge isn't enough. Most people drop out because they learn Claude in a vacuum. Build around real weekly tasks, and it can be internalized in a few days. Mikky added that the third setting is easy to write, but most people can't be bothered to write it. These points aren't new, but next to this checklist they become actionable.

Zephyr's original article: https://x.com/i/article/2055228165920874496
Original tweet: https://x.com/Zephyr_hg/status/2116719386245644512
One filename in the image is written as CLAUDEx.md, while the main text says CLAUDE.md; just understand it from context.