Judge
Checks a post, article, deck, proposal or landing page before it goes out. A script catches what can be counted, then a fresh subagent that never saw your brief scores it out of 100. The bar is 90.
For anyone who wants an outside score before publishing. Free. No terminal: ZIP · 21 KB
npx skills add victor-shulga/judge/skills/judgeWhat it does
Whoever wrote it does not grade it. A script goes through the work first: banned names, links, numbers, AI-sounding phrases. Then Claude starts a separate subagent that sees only the work itself, with no brief and no earlier drafts. It scores 30 points on shared criteria and 70 on the criteria for that type. Below 90 or with a hard fail, you get one concrete fix and a new judge scores the next round, three rounds at most. The skill never edits your work. What to do below the bar stays your call. First run on my own /stack/ page: 66, 74, 78 over three rounds, it never reached 90.
Input and output
Key features
- Blind scoring: a fresh subagent every round, with no brief and no earlier drafts
- 8 types of work: post, article, deck, proposal, landing page, site, creative, email sequence
- Every download link is checked by content type: a page served instead of a file counts as an error
- Every number is listed for checking, and the script catches "70 skills" here and "69" there
- Your banned list (clients, vendors) stays on your machine and is never part of the skill
- Needs Claude Code and Python 3. For posts add content-engine-skills, for AI-sounding phrases add anticopywriting-ai; without them the skill still runs and says what it skipped
How to install
Node 18+ required. The command installs the skill into the current project folder:
npx skills add victor-shulga/judge/skills/judgeTo keep the skill at hand in every project, add -g:
npx skills add -g victor-shulga/judge/skills/judgeIn Claude Code: unzip the folder judge into ~/.claude/skills/ and restart the app. The skill runs only in Claude Code: it needs Python 3 and a separate subagent.
Download the ZIP · 21 KBOpen the skill's folder on GitHub and save the files. The main file is SKILL.md, and the rest is pulled in from it.
Check it works: type /judge in the chat — the skill should answer. If it does not: skills are read at session start, so restart Claude.