# The Judgment Folder Starter Kit From "Applied AI for Marketing Ops" (appliedaiformops.com) by Lily Luo. This is the system I use to stop making the same corrections twice: a folder that captures how I decide things, in a format AI tools read before I see a draft. Mine picked up about 30 rulings in its first two days and my editing sessions have been getting shorter since. **Pick your level:** - **Level 1 — Manual.** One document, this week, no tools. - **Level 2 — The full setup.** For Claude Code or Cowork users: a git-tracked folder with a feedback loop that turns your reactions into standing rules. - **Level 3 — Hand it to your agent.** Paste this whole file to your AI agent and say "set this up for me." The last section is instructions for the agent. The problem this solves: every time you correct an AI draft, that correction disappears into the chat history. Next draft, same corrections. Your judgment never compounds. --- ## Level 1: The rulings document (one document, this week) ### Step 1. Open a document titled "Rulings" Four columns, or just lines in this order: **date · what I corrected · my exact words · the rule behind it**. ### Step 2. Log corrections as you make them Next time you fix an AI draft, take 20 seconds to log it. The "my exact words" column matters more than it looks; your phrasing carries the judgment. Real examples from my log: ``` 2026-07-10 · em dashes everywhere · "there are too many em dashes here" · Max one em dash per page. Prefer commas and periods. 2026-07-10 · hype phrasing · "remove typical LLM speak" · Ban: delve, seamless, game-changer, "it's not X, it's Y." 2026-07-11 · off-brand colors in a mockup · "this is too Claude-like" · Use only the approved palette. Never propose new brand colors. ``` Only log corrections you'd make again. A one-off edit ("fix this typo") is not a ruling. A repeatable preference ("never open with a rhetorical question") is. ### Step 3. Paste the file into your next prompt Before the AI drafts anything, attach the Rulings doc with one line: "Follow every rule in this document. Before returning your draft, list any rule you were at risk of breaking." Watch how many of your usual edits disappear. ### Step 4. Track the misses Add a "Misses" section: every time a mistake you already corrected shows up again, log it. This is the number that matters. If misses stay near zero, your corrections are compounding. My metric: each editing session should be shorter than the last for the same kind of work. --- ## Level 2: The full setup (Claude Code or Cowork) This is how mine actually runs: a small git repository (version control gives you history and portability; any folder works to start) with three layers, and a documented loop that keeps it current. ### The folder ``` judgment/ voice.md hard rules about how you write and speak positioning.md what you do and don't say publicly design.md settled taste: approved palettes, formats, structures rulings-log.md the append-only log from Level 1, with a Misses section strategy/ philosophies.md your core beliefs, distilled from things you've actually published or presented, with quotable lines routines/ feedback-compiler.md the written procedure for the loop below ``` The routines folder matters more than it looks. The loop itself is documented as a file, so any agent, or any future version of you, can read exactly how it's supposed to run. ### Seeding it: extract, don't introspect Don't write your rules from memory; you'll write who you'd like to be, not how you actually decide. Instead, point Claude Code at the evidence: past drafts with your tracked changes, old feedback you've written, your published work. My instruction was a version of: > Read these documents. Extract the corrections I make repeatedly and the patterns in how I decide. Propose each one as a candidate rule, with the evidence quoted. I'll approve or reject each. Approved rules land in the judgment layer. Rejected ones get logged too; knowing what you're NOT is judgment. ### The loop that keeps it current This is the part that compounds. Every time you react to a draft (I usually dictate my reactions; typed comments work the same): 1. The agent splits your feedback into two piles: **one-off edits** (applied now, not logged) and **candidate rules** (generalizable). 2. Candidate rules come back to you as a short approve/reject list. 3. Approved rules land in the right file, with a dated entry in `rulings-log.md` quoting your exact words. 4. Every future draft gets checked against the judgment layer before you see it. A practical note: the loop lives or dies on capture friction. If logging a reaction takes longer than making the edit, you'll stop logging. Voice capture fixed this for me; a dictation tool that drops your spoken reactions into the loop makes rulings nearly free. ### The two numbers to watch Misses (mistakes that were already ruled on showing up again; should stay near zero) and session length (each editing session shorter than the last for the same kind of work). If either goes the wrong way, the loop needs attention, not more rules. --- ## Level 3: For your AI agent *Human: paste this whole file to your agent and say "set this up for me." Agent: follow this section literally.* 1. Ask the human one question first: single rulings document (Level 1) or the folder version (Level 2)? Build only what they choose. 2. **Seed from history, not from scratch.** Ask for past examples of the human's edits and feedback. Extract repeated corrections, propose each as a candidate rule with evidence quoted. Only approved rules get written. 3. **Split every piece of feedback into two piles**: one-off edits (apply now, don't log) and candidate rules (propose for approval). Never promote a correction to a rule without approval. 4. **Log rulings with the date and the human's exact words.** Paraphrasing loses the judgment. 5. **Check every draft against the judgment folder before the human sees it**, and say which rules you applied. 6. **Maintain the Misses section.** When the human corrects something the folder already covers, log the miss and ask whether the rule needs sharpening. 7. This folder is the human's accumulated judgment. Never delete or rewrite entries without explicit approval; append instead. --- Questions, or a folder that's saving you real editing time? Contact lily@lilyluo.ai.