← ClaudeAtlas

tellslisted

Audit finished work, prose or a built interface, for the patterns that give machine-made output away. Reads one shared register, quotes the offending line or file:line for each check, then fixes what fails.
afterlaunch/skills · ★ 1 · AI & Automation · score 75
Install: claude install-skill afterlaunch/skills
# Tells An audit, run **after** the work exists. Never a brief handed to whatever is doing the writing or building. That placement is not a style preference. Structure belongs at the boundaries of a piece of work, not inside the act of making it: assemble what you need before you start, check what you produced after you finish, and put nothing in between. A model, or a person, handed a coverage checklist while it writes tends to satisfy the checklist and lose the actual point; a register run afterwards catches the same problems without causing that one. Read `TELLS.md` first. It is the single source; do not work from memory of it. Then read the words this business refuses. They live in `product-context/PRODUCT.md` in the estate (`~/.claude/content/product-context/PRODUCT.md` by default), in the refused words list and the never-say list. Each entry is a check in this run, exactly like a register entry: quote the line it turns up on and fix it. A refused word is a defect even where the register would have let it through, because the register is generic and that list is yours. If the file is not there, say so once in the report: the audit is running without half its vocabulary. ## What this catches that a green build does not A build can enforce the mechanical subset of this register (Part 3 is written as literal patterns for exactly that). This skill exists for the rest: the judgement calls a regex cannot make. - Does this grid have real visual variety, or six identical