Guide · Agent skill
How to use the Kill Argument Adversarial Review agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The kill-argument skill is an agent skill for Claude Code and compatible agents that stress-tests a paper's headline claims. It has a fresh reviewer write one roughly 200-word rejection argument, then a second fresh reviewer defends the paper point by point. After this guide you will know what the skill does, how to install it, how to trigger it with a natural request, and what it needs to run.
Kill Argument Adversarial Reviewwanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Download and unzip. Download the skill ZIP from this page and unzip it into ~/.claude/skills/ or the project's .claude/skills/ folder.
- 2Or upload in the Claude apps. In the Claude apps, go to Settings → Capabilities → Skills and upload the skill.
- 3Confirm the reviewer is available. Make sure a Codex MCP server or another configured reviewer tool is set up, and that you have access to the reviewer model.
- 4Ask for the review. Say something like "Run a kill argument review on paper/", pointing at your paper folder.
- 5Wait for the attack and defense. The skill runs a fresh attack thread, then a fresh defense thread that classifies each objection.
- 6Read KILL_ARGUMENT.md. Open the report and start with the still_unresolved items to decide what to revise.
What is an agent skill, and what does kill-argument make the agent do?
An agent skill is a folder containing a SKILL.md file that tells an agent how to carry out a specific task. kill-argument is one such skill, aimed at theory papers that have already been through standard review rounds.
Standard score-based reviews tend to produce balanced weakness lists. The kill-argument skill instead asks one reviewer to commit to the single most damaging rejection argument, the kind of sentence a senior area chair could use to reject the paper. A second fresh reviewer then defends the paper against each objection and sorts each one as answered by current text, partially answered, or still unresolved.
- Attack stage: a fresh reviewer writes a single argument of about 200 words aimed at the headline claims.
- Defense stage: a second fresh reviewer splits the attack into 3-7 atomic points and classifies each one.
- Outputs: KILL_ARGUMENT.md and KILL_ARGUMENT.json in the paper directory, and an HTML version may be generated.
Install the kill-argument skill
Download the ZIP from this page and unzip it. You can install it for yourself across all projects, or for a single project.
For a personal install, unzip into ~/.claude/skills/. For a project install, unzip into the project's .claude/skills/ folder. Alternatively, in the Claude apps, open Settings → Capabilities → Skills and upload the skill there.
- Personal: ~/.claude/skills/
- Project: .claude/skills/ inside your project folder
- Claude apps: Settings → Capabilities → Skills, then upload
Set up the reviewer backend
By default the skill uses the Codex MCP server (mcp__codex__codex) as its reviewer. Other reviewer tools named in the skill (antigravity, grok, claude-review) are allowed if you have configured them. You can choose a reviewer by adding an optional .aris/reviewer.json file in your project.
The skill names gpt-6-astra as the reviewer model, with fallbacks. Using it requires access to the reviewer model, which implies an OpenAI-compatible API account.
- Requires Claude Code or a compatible agent with the skill installed.
- Requires a Codex MCP server or another configured reviewer tool.
- Optional: .aris/reviewer.json to select the reviewer.
Trigger the skill with a natural request
You do not need to remember a command. The skill's description lists phrases that activate it. Ask for the review in plain language and point the agent at your paper folder, which the skill takes as its argument.
The skill is meant for theory papers that have already passed standard review rounds, shortly before submission or during rebuttal preparation.
- "Run a kill argument review on paper/"
- "Do an adversarial review of my paper folder before I submit"
- "Simulate reviewer-2 for this theory paper"
- "Prepare a rebuttal: hostile review of paper-overleaf/"
Read the results and know the limits
The skill writes KILL_ARGUMENT.md for reading and KILL_ARGUMENT.json for machine use. Each objection is classified as answered_by_current_text, partially_answered, or still_unresolved. Focus first on the still_unresolved items, since they are the points most likely to survive a real review.
The verdict is generated by language models and can be wrong or overly harsh. Treat it as a prompt for checking your framing, not as ground truth. The skill is designed to run once per draft; repeated runs add little unless the paper changed.
- Do not wrap it in /loop, /schedule, or CronCreate; it is verdict-bearing and designed to run once.
- Not suited to empirical papers without theorems; use /research-review for those.
- Skip the HTML step by passing the option to set render html to false if you do not need it.
Safety notes: scripts and network access
This skill runs scripts and uses the network. Its allowed tools include shell commands and calls to reviewer MCP tools, and the reviewer runs on a remote model. Your paper text is sent to the configured reviewer. Check your reviewer configuration and what your account permits before running it on unpublished work.
- Runs scripts: yes (shell commands for locating files in the paper directory).
- Uses network: yes (reviewer calls to the configured model).
- The skill is licensed under MIT.
What people use it for
- Finding framing overclaims, such as a setting that is mostly conditional rather than general, before submission.
- Predicting reviewer-2's strongest objection to prepare a rebuttal in advance.
- Checking whether a high-level title oversimplifies the actual theorem.
- Surfacing scope or assumption-versus-claim mismatches that score-based reviews did not flag.
Frequently asked questions
Is the kill-argument skill free to use?
The skill is released under the MIT licence. The reviewer runs on a configured model, which implies an OpenAI-compatible API account, so model access may carry its own cost.
What do I need to run kill-argument?
You need Claude Code or a compatible agent with the skill installed, a Codex MCP server as the default reviewer backend, and access to the reviewer model named in the skill. An optional .aris/reviewer.json file lets you choose a different reviewer.
Where do I install the kill-argument skill?
Unzip the downloaded ZIP into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. In the Claude apps you can instead upload it under Settings → Capabilities → Skills.
When should I use kill-argument instead of research-review?
Use kill-argument for theory papers that have already passed standard review rounds and are close to submission. For empirical papers without theorems, the skill's own guidance points to /research-review instead.
Can I run kill-argument on a schedule?
No. The skill is verdict-bearing and designed to run once. Re-running on a timer adds no new signal unless the paper changed, so schedule the work that comes before it, such as stabilising the draft, and then run it once.
Can I trust the verdict?
Not as ground truth. The verdict is generated by language models and can be wrong or overly harsh, so use it to check your framing and scope claims.
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
Use kill-argument if you are a researcher or author with a theory paper that has passed standard review and you want to anticipate the hardest rejection argument before submission or rebuttal. Pick something else for empirical papers without theorems, for repeated automated reviews, or if you cannot use a remote reviewer model on your unpublished work.
Download the skill