The Antagonist
A free Claude Skill that plays the hard critic: it finds what is wrong in your code, plans and texts, proves each point, and changes its mind only for evidence.
Version 1.0 ·
Download the skill
Free, no account, no sign-up. A zip file with SKILL.md, three reference files and the licence.
Download (English) 10 KB
Édition française · 11 KB
Why a critic, and why a skill
AI assistants lean toward agreeing with whoever is typing. They open with praise, soften every problem and drop a point as soon as you push back. That feels good and costs you later: the bug ships, the plan meets reality, the article goes out with a statistic nobody checked.
Asking "be brutally honest" in a chat helps for one answer. The Antagonist makes it stick. It is a Claude Skill: a folder of instructions Claude loads whenever you ask for a review, so the same rules apply every time, in every project.
It is not built to disagree with everything. A critic that objects to everything is as useless as one that agrees to everything. Its job is to look at your work with a new set of eyes, and to prove each objection.
The seven rules it follows
- Restate before attacking. It first states your work in its strongest form, so it attacks what you meant, not a misreading.
- Every objection carries proof. Severity (Fatal, Serious, Minor), confidence, the evidence (a line, a quote, a number) and a test that would settle it. No evidence, no objection.
- No praise, no padding. What holds up gets one line at the end.
- A clean result has to be earned. "No serious problem" comes with the list of what was checked.
- Questions before the verdict. It asks you the three hardest questions first, so you do the thinking too.
- It holds its ground for the right reasons. It changes its mind for evidence or a better argument, and says what changed it. It does not change it because you insist.
- It says what it could not check. Code it could not run, data it did not see, facts it could not verify.
Three modes
| Mode | Use it on | What it looks for |
|---|---|---|
| Code | Code, pull requests, scripts, configs | It runs what it can, then hunts edge cases, silent failures, security holes, invented functions and tests that cannot fail |
| Plan | Plans, strategies, ideas, decisions | Unchecked assumptions, a pre-mortem, the cheaper path, the real cost, who loses, and the signal to stop |
| Slop | Articles, posts, emails, reports, pages | Empty phrases, unsourced numbers, invented quotes and sources, vague claims, structure on autopilot |
Every review ends the same way: a count (Fatal, Serious, Minor, Withdrawn) and one verdict: ship, fix first or rethink.
How to install it
| Where you use Claude | What to do |
|---|---|
| Claude Code | Unzip the file into ~/.claude/skills/ (for you) or .claude/skills/ inside a project (for the whole team) |
| Claude apps and Claude.ai | Turn on Skills in Settings (Pro, Max, Team, Enterprise) and upload the zip file |
| Other agents | Agents that read SKILL.md files can use the same folder; check your tool's documentation |
How to use it
Ask for a review in plain words: "tear this apart", "play devil's advocate on this plan", "what's wrong with this page?", or type /antagonist in Claude Code. Then argue back. If you show it new evidence, it will withdraw the point and say why. If you only insist, it will not.
Try this: take something you finished this week and believe is good. Run the Antagonist on it, and answer its three questions in writing before you read the objections.
What it will not do
- It will not tell you who or what wrote a text. It judges the lines, not the author.
- It does not replace a security audit, a lawyer or an accountant. It can tell you when you need one.
- It cannot check what it cannot see. Give it the files, data and context, or read its "Not checked" list carefully.
The skill is shared under CC BY-NC-SA 4.0: keep the credit and the same licence; commercial use needs written permission. How to reuse it.