Founders who want honest feedback on long posts and essays
No metric reported. Packy McCormick found Claude's grades drifted toward A-minus whatever he sent, though its notes (a section drags, writing is too defensive) were often right. On Every's AI & I, he and Dan Shipper built a rubric-based grader in a Claude Project and made it write reasoning before each score, which produced more varied grades. They concluded diverse examples beyond your own writing help most.
Shipper's instruction added to the grader: output the reasoning first for each subsection and the full grade, then the score. His explanation: the model 'thinks by writing,' so scoring first makes it rationalize the score.
Project instructions: I'm [name], I write [publication/posts] for [audience]. You are a demanding editor. When I ask you to grade a draft, use the grading rubric in the project files and compare against the example essays. For each rubric category: first write your reasoning with specific quotes from my draft, then give the score. Only after all categories, give the total. Then list: - The 3 changes that would raise the score most, in order. - Any section that drags (quote where it starts and ends). - Anywhere I hedge or get defensive without need. Do not rewrite the draft. Here's my draft. Please grade it. [paste]
Legacy Builder Lab is Clifton Sellers' community for founders who write their own content with AI. Members get his prompts and bot setups, five new post templates every Monday and a monthly live teardown.
Join the Lab waitlist