How to use the Interview Cheat Sheet Writer agent skill

What is an agent skill, and what does this one make the agent do?

An agent skill is a folder with a SKILL.md file that tells an agent how to handle a specific kind of task. When a request matches the skill's description, the agent follows its workflow instead of a generic approach.

This skill follows a fixed template to produce one long-form Chinese tutorial per run, typically 600 lines at the default effort and about 1000 at max effort. The template has a TL;DR, intuition, core formulas with derivations, a from-scratch PyTorch implementation of 50 to 80 lines, variants and engineering notes, complexity, comparison with related methods, and 25 interview questions graded L1 (must-know), L2 (advanced) and L3 (top-lab level), with collapsible answers.

• Output is Chinese-first, with English technical terms kept in place.
• The draft is written to docs/tutorials/ as a Markdown file.
• Before rendering, the draft goes to a second AI model (via the Codex MCP server) for math, code and citation review.
• It does not commit files to git unless you approve.

How do you install the interview-cheatsheet skill?

There are two routes. For Claude Code, download the skill ZIP from this page and unzip it into your user skills folder, ~/.claude/skills/, or into a project's own .claude/skills/ folder if you want the skill only in that project. For the Claude apps, upload the ZIP under Settings, then Capabilities, then Skills.

After installing, start a new session so the agent can see the skill. The skill's files are a single SKILL.md, so you should see that file inside the unzipped folder.

• Claude Code: unzip into ~/.claude/skills/ (all projects) or <project>/.claude/skills/ (one project).
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP.

How do you trigger the skill with a natural request?

You do not need to type a command. The skill's description lists phrases that should activate it, and you can phrase requests the same way. Name a specific topic, not a broad field.

Example phrases from the skill description include '写面试 cheat sheet', '写一份 X 教程', '帮我准备 Y 面试题' and '出一份 X 速查'. You can also ask directly for a 600 to 1000 line Chinese tutorial on a specific ML topic.

The skill also documents an argument form for explicit control: a topic plus optional --effort (balanced or max), --byline and --commit flags.

• Good topics: 'RLHF / DPO / PPO', 'MoE', 'KV Cache + Speculative Decoding'.
• Too broad: 'all of LLM training' or 'diffusion'. The skill says to split these, and it may stop and ask you to narrow the scope.
• --effort balanced is the default (about 600 lines); --effort max targets about 1000 lines with deeper proofs and more L3 questions.
• --commit defaults to false, so the skill stops after rendering and you review the result yourself.

What do you need before you run it?

The skill is written for Claude Code or a compatible agent environment. Its workflow calls a cross-model review step through the Codex MCP server (mcp__codex__codex), so you need that server available. The source notes that other reviewer backends are optional and come from a shared reviewer configuration the file references.

The workflow also refers to a render step (/render-html) that is not included in this SKILL.md file. Plan on handling rendering separately. The skill needs Bash and file write permission to docs/tutorials/.

• Claude Code or a compatible agent environment.
• Codex MCP server for the review step (optional alternative backends per the referenced reviewer config).
• A /render-html step, which is not included in this file.
• Bash and file write access to docs/tutorials/.

What should you check in the safety notes?

The skill does not run scripts, according to its listing. It does use the network, because the review step sends the draft to a second model through the Codex MCP server. The skill's allowed tools include Bash, Read, Write and Edit, so review what the agent runs and writes, especially if you have granted broad permissions.

Treat the output as a draft. The cross-model review is an automated check and can still miss errors in formulas, citations or code. Verify derivations, paper attributions and runnable code yourself before studying from it or publishing it.

• Scripts: none listed.
• Network: yes, for the cross-model review.
• Verify formulas, citations and code independently.
• Keep --commit at false unless you have reviewed the output.

What are good tips for getting useful output?

Pick one technique or a tightly bounded pair, such as RLHF versus DPO, instead of a whole field. Narrow topics produce tighter formulas and more focused interview questions.

Read the draft's comparison tables and L3 questions with extra care, since those are where subtle errors are most costly. Check the code blocks by running them, because the skill's own review is automated.

• Name the exact method and its common alternatives.
• Use --effort max when you want deeper derivations and more L3 questions.
• Set --byline only if you plan to render the piece with an author line.
• Read the output in Chinese or use a translation step; the template is Chinese-first.