Guide · Agent skill
How to use the Anti-Sycophancy Check agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The anti-sycophancy skill is an agent skill that changes how an AI assistant responds when you bring it a claim, plan or piece of code. Instead of agreeing with your framing, the assistant restates your claim in plain terms, checks it against the evidence and gives its conclusion before its reasoning. This guide covers what the skill does, how to install it from the ZIP on our page, how to trigger it, and where it falls short.
Anti-Sycophancy Checksickn33/agentic-awesome-skills · MIT Download the skill Quick start
- 1Download the ZIP. Get the anti-sycophancy ZIP from our page and unzip it on your computer.
- 2Place the folder. Move the anti-sycophancy folder into ~/.claude/skills/, or into .claude/skills/ inside a project.
- 3Or upload in the Claude apps. In the Claude apps, open Settings → Capabilities → Skills and upload the skill there.
- 4Load the skill. Type /skill anti-sycophancy at the start of your session.
- 5Make a specific request. Paste the claim, plan or code you want checked, and ask the agent to review it with the skill active.
- 6Check the conclusion. Read the conclusion first, then the evidence, and confirm the evidence against your own sources.
- 7Push back with evidence. If you disagree, add new evidence rather than repeating your view, so the agent can decide whether to update.
What does the anti-sycophancy skill make the agent do?
Sycophancy means an AI assistant leaning toward agreeing with you. The SKILL.md file describes a fixed process the agent follows for every response while the skill is active. It first extracts your core claim and states it in one sentence without the premises. It then assesses that claim on its own, looking at evidence for and against, without referring to whether you agree or who you are. Its conclusion comes from that assessment alone, and the response puts the conclusion first and the evidence second.
The skill also covers pushback. If you disagree, the agent sorts your reply into new evidence or repeated opinion. New evidence leads it to update its position and say what changed. A repeated opinion leads it to restate its position with the evidence.
- Restates your claim in one plain sentence
- Assesses the claim on evidence, not on your authority or framing
- Gives the conclusion first, then the evidence
- When you push back, changes position only for new evidence
How do you install the anti-sycophancy skill?
Download the ZIP from our page and unzip it. Put the anti-sycophancy folder in ~/.claude/skills/ for your user account, or in a project's .claude/skills/ folder to limit it to that project. In the Claude apps you can instead upload the skill under Settings → Capabilities → Skills.
The source README is written for OpenCode and gives a different location, ~/.config/opencode/skills. If you use OpenCode, copy the folder there with the commands below.
- Claude Code, user-wide: ~/.claude/skills/
- Claude Code, one project: .claude/skills/ inside the project
- Claude apps: Settings → Capabilities → Skills, then upload
mkdir -p ~/.config/opencode/skills
cp -r anti-sycophancy ~/.config/opencode/skills/How do you trigger the skill with a request?
You can load the skill explicitly by typing /skill anti-sycophancy. Once it is loaded, the agent applies the process to every response in that session. You can also name the skill inline in a normal request, as in the example below. Explicit loading is the more reliable option when you want the agent to stay direct for a whole conversation.
Use this skill when you want a second opinion on a decision, not help carrying one out. Ask it to review a design, challenge an estimate, or check whether an approach holds up.
- Load it for a session: /skill anti-sycophancy
- Name it in a request: ask the agent to review something with the skill active
- Ask directly for a challenge: state your plan and ask what is wrong with it
/skill anti-sycophancyWhat do example requests look like?
The source README gives this inline pattern. Replace the last part with your own material, such as a design doc, a function or a plan.
Good requests give the agent something concrete to assess. A vague request like 'is this good?' gives it little to test. Pasting the claim and the context you are relying on gives it a claim to check, and gives you an answer you can verify.
- With the anti-sycophancy skill active, review this architecture: [paste your design]
- With the anti-sycophancy skill active, I think caching every query will fix our latency. Check that claim against the code.
- Push back with evidence, e.g. a benchmark you ran, and watch the agent update its position and say what changed
With the anti-sycophancy skill active, review this architecture:What does the skill need, and is it safe to run?
The listing states that the skill requires nothing beyond the agent itself. It does not run scripts and does not use the network. The package contains README.md and SKILL.md. Because it is only instructions, the main thing to check is whether you want the assistant's posture to change for the session.
The skill changes how the assistant responds, not what information it can reach. It does not add data sources. Questions that need current or external facts still depend on the assistant's own tools and sources.
- Requires: nothing listed
- Runs scripts: no
- Uses network: no
- Adds data sources: no
What are the limits and tips for better results?
The skill tells the agent not to be reflexively contrarian when a claim is already supported by evidence. Even so, responses can be more confrontational than you need when you only want a task done. Turn it off by not loading it for ordinary execution work.
Ask for the evidence you need. The skill makes the assistant state its basis, so ask it to name the files, numbers or sources it used.
- Do not use it for simple task execution; responses can be more confrontational than needed
- Expect no new facts: claims still need evidence from the code, tools or sources available
- Support your pushback with new evidence, since repeated opinion will not change the answer
What people use it for
- Getting a second opinion on a technical architecture before committing to it
- Challenging an estimate or plan that someone has already approved
- Reviewing code for claims that a change is safe or faster without evidence
- Testing your own assumptions before presenting them to a team
Frequently asked questions
How do I install the anti-sycophancy skill?
Download the ZIP from our page and unzip it. Place the anti-sycophancy folder in ~/.claude/skills/ or in a project's .claude/skills/ folder, or upload it in the Claude apps under Settings → Capabilities → Skills.
How do I turn on the anti-sycophancy skill?
Type /skill anti-sycophancy in your session. Once loaded, the agent applies its process to every response. You can also name the skill in a request.
Is the anti-sycophancy skill free?
The skill is licensed under MIT, as stated on the listing. Check the license file in the source repository for the full terms.
What does the skill require to run?
The listing says it requires nothing beyond the agent itself. It runs no scripts and uses no network connection.
Does the skill make the AI more accurate?
No. It changes the assistant's response posture, not its access to facts. Claims still need evidence from the code, tools or sources the assistant has.
Why is my AI still agreeing with me?
Check that the skill is loaded in the current session. If you pushed back without new evidence, the agent is designed to restate its position rather than change it.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
This skill suits developers and reviewers who want claims challenged on evidence, such as architecture reviews and plan checks. Skip it for routine task execution or for factual questions that need current external data, since it does not add sources and can make replies more confrontational than needed.
Download the skill