Guide · Agent skill
How to use the Cheat on Content agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
Cheat on Content is an agent skill for content creators. It helps you turn gut feelings about what will perform into predictions you can check against real results. This guide explains what the skill makes an agent do, how to install it, how to start it with natural requests, and what it needs and touches on your machine.
Cheat on ContentXBuilderLAB/cheat-on-content · MIT Download the skill Quick start
- 1Download and unzip. Download the ZIP from this page and unzip it. Read install.sh and uninstall.sh first.
- 2Place the skill folder. Put the cheat-on-content folder in ~/.claude/skills/ for all projects, or in your project's .claude/skills/ folder. In the Claude apps, upload the ZIP under Settings → Capabilities → Skills instead.
- 3Open your content project. Change to your content project folder in Claude Code or Codex so the skill can create and read project files there.
- 4Run init. Run /cheat-init in Claude Code, or say 初始化 cheat-on-content in any agent. Approve the files it proposes to create.
- 5Add benchmark accounts. Say 找对标 or learn from to import competitor accounts as initial signal sources. This is strongly recommended for new users.
- 6Score and predict a draft. Say 启动预测 with your draft path, such as scripts/foo.md, to get a scored draft and a locked prediction before publishing.
- 7Record results and review. After publishing, tell the agent it shipped, then ask for a 复盘 (retro) once the review window has passed.
What does the Cheat on Content skill make the agent do?
The skill runs a loop: score a draft against a rubric, write a prediction before any results exist, publish, review actual performance against that prediction, and then update the scoring rules. The method is meant to work for any content you can measure, such as views, reads, listens or clicks.
The rubric is the part you can swap out. The built-in rubric is for short-form opinion videos (commentary, current-affairs takes, argument pieces and personal opinion posts), with seven dimensions. If you make other formats, the README says you should write your own rubric, using the starter-rubrics/opinion-video-zero.md file as a model.
- Scores a draft on the rubric dimensions, with reasons and comparison anchors
- Locks a blind prediction in a log before publishing
- Reviews performance after a set window and records what happened
- Proposes rubric upgrades, which must re-score past samples and pass an independent check
- Helps with competitor study, topic ideas and trend tracking
How do you install the Cheat on Content skill?
Download the ZIP from the skill's page on this directory and unzip it. For Claude Code, place the cheat-on-content folder in your user skills folder (~/.claude/skills/) or in a project's .claude/skills/ folder. You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
The package includes install.sh and uninstall.sh scripts. Read them before running anything, because the skill runs scripts on your machine.
- Claude Code: unzip into ~/.claude/skills/ for all projects, or into <your-project>/.claude/skills/ for one project
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
- Codex: the skill works through natural-language routing, so you do not need slash commands
~/.claude/skills/cheat-on-content/How do you start a project with /cheat-init?
The first time you use the skill, you must run the init step in your content project folder. Init creates the project files the skill expects, such as the rubric notes, a workflow document, a status board and a state file. It asks before overwriting anything that already exists.
In Claude Code, the command is /cheat-init. In Codex or any other agent, say 初始化 cheat-on-content ("initialise cheat-on-content").
- Init creates a .gitignore that keeps credential files out of version control
- Init creates a predictions/ folder, which is protected by a hook that stops edits to locked predictions
- Setting up benchmark accounts with /cheat-learn-from is strongly recommended for a new user
/cheat-initHow do you trigger the skill with natural requests?
You do not have to memorise commands. The skill maps common phrases to sub-skills. In Claude Code you can use the slash command, and in any agent you can use a natural request with the same meaning. Use the file path of your draft when scoring or predicting.
The table below lists the phrases the skill recognises, taken from its description and routing table.
- 打分这篇 scripts/foo.md (score this draft) → scoring only, nothing is written to a file
- 启动预测 scripts/foo.md (start prediction) → scores the final draft and writes a locked prediction
- 拍了 / 录完了 (shot it) → records that the video was filmed
- 已发布 (I shipped it) with the publish link → records publication
- 复盘 (retro this) → reviews results once the review window has passed
- 找对标 / learn from → imports benchmark accounts
- 推荐选题 (next topic) → ranks candidate topics
- 抓热点 (fetch trends) → pulls trend ideas into the candidate pool
- 状态 (status) → shows the dashboard, including buffer warnings
- 升级 rubric (bump rubric) → proposes a rubric upgrade
What does a typical workflow look like?
A typical cycle starts with a draft in your scripts/ folder. You ask for a score, then ask for a prediction. After you film and publish, you tell the agent what happened and later ask for a retro. The skill shows each dimension score and at least one comparison anchor. It will not pick a topic with only a single composite number.
The skill refuses some requests by design. It will not change a locked prediction, back-fill results you have not collected, skip the rubric re-score on an upgrade, or lower the upgrade threshold to get a change through.
- Describe a draft and get a score with reasons, not a verdict with no explanation
- Predictions cannot be rewritten; a justified redo is saved as a new _redo file and the original stays
- Rubric observations that new data contradicts are deleted, because git history is the archive
What do you need, and what does the skill touch on your machine?
The skill needs Claude Code or Codex with the skill installed, plus Bash and filesystem access in your content project folder. It runs scripts and uses the network, so review the package before you install it.
Some features are optional. A cross-model LLM chat tool (mcp__llm-chat__chat) can provide an independent audit during rubric upgrades. Trend adapters for the trends feature may need external accounts or APIs; the reviewed files do not specify which ones. The skill has a hook that enforces locked predictions in Claude Code. Credential files are listed in the generated .gitignore, so keep your own secrets out of shared folders.
- Runs scripts: yes. Uses network: yes
- Optional: cross-model audit via mcp__llm-chat__chat. Cross-model audit can only be turned off by explicitly setting CROSS_MODEL_AUDIT=false in the state file
- Optional: trend-source adapters, which may need accounts or APIs you must set up yourself
- Review install.sh and uninstall.sh before running them
What tips help you get value from it?
The value comes from repeating the loop. One prediction tells you little; a calibrated model needs several published samples. The skill has two modes: cold-start for new creators, which uses a simpler prediction with seven dimension scores and a one-sentence bet, and calibration for creators with five or more reviewed posts, which unlocks the full prediction.
Keep your real numbers. The skill is built to refuse to invent or back-fill results, so record actual data when you have it.
- Set up benchmark accounts early so the rubric has real signals to start from
- Predict before you look at any numbers, and publish on the schedule you set
- Use the status command regularly to check buffer warnings and pending reviews
- Do not use the built-in rubric for article, podcast or other formats without adapting it
What people use it for
- Checking a first opinion-video draft against a structured rubric before filming
- Comparing a planned post with benchmark accounts to see which patterns it matches
- Tracking whether your predictions match real performance over several published videos
- Ranking candidate topics by which ones fit your current content buffer and mix of safe and experimental ideas
Frequently asked questions
Is the Cheat on Content skill free, and what is its licence?
The skill is released under the MIT licence, as stated in its listing. The reviewed files do not mention prices, so check the page for any cost details before you install.
What do I need to run the Cheat on Content skill?
You need Claude Code or Codex with the skill installed, plus Bash and filesystem access in your content project folder. An optional cross-model LLM chat tool supports independent audits, and some trend features may need external accounts or APIs.
How do I install the Cheat on Content skill in Claude?
Download the ZIP, then either unzip the folder into ~/.claude/skills/ or a project's .claude/skills/ directory, or upload the ZIP in the Claude apps under Settings → Capabilities → Skills. Read the install.sh script first, since the skill runs scripts.
Does Cheat on Content work for articles, podcasts or newsletters?
The method is meant to apply to any measurable content, but the built-in rubric is for short-form opinion videos. For other formats, you need to write your own rubric, using the opinion-video starter as a template.
Can Cheat on Content back-fill or invent my video results?
No. The skill refuses requests to back-fill or fabricate results, and it will not let you edit a locked prediction. A justified redo is saved as a separate file so the original stays in place.
How is this different from asking an AI for a quick critique?
A one-off critique gives feedback once. This skill keeps a log of blind predictions, reviews them against real performance, and refines the scoring model over repeated cycles, so its value builds over time.
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 Cheat on Content if you make short-form opinion videos or similar content, want to check whether your sense of what will perform is accurate, and are willing to keep recording real results over several posts. Pick something else if you want a one-time critique, need a ready-made rubric for articles or other formats, or are not comfortable with a skill that runs scripts and uses the network.
Download the skill