Guide · Agent skill
How to use the Cross-Model Memo Review agent skill (cross-eval)
Updated · written with AI from the project's own documentation and checked against our catalogue data
The cross-eval skill is a set of instructions that tells an AI agent, such as Claude Code, to run a high-stakes business memo past several different AI models and compare their verdicts. It is aimed at founders and executives who are about to make an irreversible decision, such as a fundraise term sheet, a layoff plan or a bet-the-company pivot. After this guide you will know what the skill does, how to install it, how to ask for it in plain language, and what to check before trusting the output.
Cross-Model Memo Reviewalirezarezvani/claude-skills · MIT Download the skill Quick start
- 1Download and unzip. Download the skill ZIP from this page and unzip it so the folder containing SKILL.md is in place.
- 2Place the skill. Move the folder into ~/.claude/skills/ for personal use, or into the project's .claude/skills/ folder, or upload it under Settings → Capabilities → Skills in the Claude apps.
- 3Connect optional providers. Make OPENAI_API_KEY or the codex CLI and GEMINI_API_KEY or the gemini CLI available if you want those models to review the memo.
- 4Prepare the memo. Have your memo or strategy brief ready as text, a file or a link the agent can read.
- 5Run the review. Use /c-level-agents:cross-eval with the memo, or ask the agent in plain language to run it through cross-eval.
- 6Read the saved report. Open the markdown file in ~/.claude/cross-eval/, check the vote tally and consensus concerns, then review the open questions.
What the cross-eval skill makes the agent do
Cross-eval asks the agent to send the same memo to up to three model providers: Claude, plus OpenAI's Codex and Google's Gemini if they are set up. Each model is asked to list its top three concerns, its top three supports, and a vote of APPROVE, REJECT or DEFER. The prompt tells the reviewers to assume the memo's reasoning is flawed until proven otherwise.
The agent then compares the reviews. Points that two or more models raise are grouped as consensus concerns, points raised by only one model are listed as divergent concerns, and the disagreements become questions for the founder. The result is saved as a markdown report.
- Claude is the primary reviewer and is always assumed to be available.
- Codex is used if OPENAI_API_KEY or the codex CLI is available.
- Gemini is used if GEMINI_API_KEY or the gemini CLI is available.
- The report includes a vote tally, consensus and divergent concerns, consensus supports, a GO / PAUSE / STOP recommendation and open questions.
Install the cross-eval skill from the ZIP
Download the ZIP for the skill from this page and unzip it. The skill is a folder that contains a SKILL.md file. Put the folder in one of two places: your personal skills folder at ~/.claude/skills/, or the .claude/skills/ folder inside a specific project if you want it only there.
If you use the Claude apps instead of a terminal, you can upload the skill under Settings → Capabilities → Skills rather than unzipping it by hand.
- Personal install: unzip into ~/.claude/skills/
- Project install: unzip into the project's .claude/skills/ folder
- Claude apps: Settings → Capabilities → Skills, then upload the skill
Set up the optional model connections
The skill works with Claude alone, but its main value is comparing several providers. To add OpenAI, make the OPENAI_API_KEY environment variable available or install the codex CLI. To add Google, make GEMINI_API_KEY available or install the gemini CLI. The skill checks which of these are present each time it runs.
The source does not give setup instructions for these keys or CLIs, so follow each provider's own documentation for those steps.
- OPENAI_API_KEY or the codex CLI enables the OpenAI review.
- GEMINI_API_KEY or the gemini CLI enables the Google review.
- The skill needs write access to ~/.claude/cross-eval/ to save the report.
Trigger the skill with a natural request
You can invoke the skill with the slash command /c-level-agents:cross-eval followed by your memo or brief. You can also ask in plain language, for example asking the agent to get a second opinion from several models on a board memo before you send it. Pass the memo itself, or a link or file the agent can read.
Use it when the decision is hard to reverse. The source lists cases such as signing a term sheet, announcing a layoff, committing to a regulated market, or any decision where reversing would cost more than six months of company time.
- Slash command: /c-level-agents:cross-eval <memo-or-brief>
- Plain request: "Run my fundraise term sheet memo through cross-eval before the board call."
- Plain request: "Get independent reviews from Claude, Codex and Gemini on this layoff plan."
/c-level-agents:cross-eval <memo-or-brief>Read the report and the recommendation
The saved report is named with the date and a slug of the memo title. Start with the vote tally, then read the consensus concerns, since those are raised by more than one model. Next, read the divergent concerns and the open questions for the founder, because a single model's objection can still be worth checking.
The recommendation follows fixed rules in the skill. GO requires 2 or more APPROVE votes and no CRITICAL concerns from any model. PAUSE applies if any model votes DEFER or any concern is CRITICAL. STOP applies if 2 or more models REJECT.
- Consensus concerns: flagged by 2 or more models.
- Divergent concerns: flagged by one model, marked as worth a second look or likely noise.
- GO routes to /c-level-agents:decide, PAUSE to /c-level-agents:freeze, STOP to re-running /c-level-agents:boardroom.
~/.claude/cross-eval/YYYY-MM-DD-<slug>.mdWhat happens if only Claude is available
If no other provider is configured, the skill runs three separate Claude passes with different system prompts: a standard reviewer, a devil's advocate that must find three critical concerns, and a steelman that must find three strongest reasons to approve. The report is labelled as single-model.
The source says this is weaker than a true multi-model review and should be treated as suggestive rather than conclusive.
- Pass 1: standard reviewer
- Pass 2: devil's advocate
- Pass 3: steelman
Safety and privacy notes
The skill does not run scripts, but it does use the network, because it sends your memo to the model providers you have connected. Think about what the memo contains before you run it, and only connect providers you are comfortable sending that content to.
The skill is not legal, tax or regulatory advice. Its own text says it is not a replacement for outside counsel or a real board. Model judgements can be wrong or overconfident, and the skill assumes that model biases differ without verifying it.
- Runs scripts: no
- Uses network: yes, to reach the connected model providers
- Requires write access to ~/.claude/cross-eval/
- Does not replace legal, tax, regulatory or board review
What people use it for
- Checking a fundraise term sheet for concerns that several reviewers raise before the board call.
- Reviewing a layoff plan for risks that a single model might overlook.
- Sanity-checking a bet-the-company strategic pivot before committing budget.
- Preparing a board memo with a list of open questions to answer before the real meeting.
Frequently asked questions
Is the cross-eval skill free?
The skill is published under the MIT licence. The source does not state prices for the model providers, so check each provider's own pricing if you connect OpenAI or Google.
What do I need to use cross-eval?
You need Claude Code or a compatible agent with the c-level-agents plugin installed, and write access to ~/.claude/cross-eval/ for the report. OPENAI_API_KEY or the codex CLI and GEMINI_API_KEY or the gemini CLI are optional for the other two reviews.
How do I install a Claude agent skill?
Unzip the skill folder into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. In the Claude apps, you can upload it under Settings → Capabilities → Skills.
Does cross-eval work with only Claude?
Yes. If only Claude is available, the skill runs three differently prompted Claude passes and labels the output as single-model. The source says this is weaker than a true multi-model comparison.
Can I use cross-eval instead of a lawyer or board?
No. The skill says it is not a replacement for outside counsel or a real board, so use it as a check before a high-stakes decision, not as sign-off.
Is it better than asking one AI model for advice?
The skill is designed around the idea that single-model recommendations have systematic biases, and that disagreement between models is useful signal. The source does not provide benchmarks showing it outperforms a single model, so judge the output on its reasoning.
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
Cross-eval suits founders or executives who have a costly, hard-to-reverse memo and already have access to several AI providers. It is a poor fit for routine decisions, for legal or regulatory sign-off, or for anyone expecting the models to be independent market experts.
Download the skill