Guide · Agent skill
How to use the Cheat Retro Review Skill agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
cheat-retro is an agent skill for Claude that handles the review step of a content prediction loop. Once a video, post or article has been live for a set number of days, it collects the performance numbers and top comments, compares them with the prediction you wrote earlier, and records what held and what did not. This guide explains what the skill does, how to install it, how to trigger it with plain requests, and what it needs before you start.
Cheat Retro Review SkillXBuilderLAB/cheat-on-content · MIT Download the skill Quick start
- 1Download and unzip. Download the cheat-retro ZIP from the page and unzip it so that SKILL.md sits inside a cheat-retro folder.
- 2Place the skill. Move the folder into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder, or upload it under Settings → Capabilities → Skills in the Claude apps.
- 3Check the project layout. Make sure your content project has prediction files, rubric_notes.md, rubric-memo.md and .cheat-state.json.
- 4Choose a data source. Use manual mode to paste metrics and top comments, or set up an adapter with its requirements and login.
- 5Trigger the review. Ask with a phrase such as "retro this" plus the prediction file path, or use the slash command with optional window and source settings.
- 6Check the output. Read the new review section and the observations written to rubric-memo.md, and confirm the prediction text was not changed.
What is an agent skill and what does cheat-retro do?
An agent skill is a folder containing a SKILL.md file that tells an AI agent how to carry out a specific routine. The agent reads these instructions when it is relevant to your request, so you do not have to re-explain the process each time.
cheat-retro covers the review phase. After a piece of content has been live for a set number of days (the default is 3), the agent pulls its results, works out ratios such as likes-to-views, checks which earlier predictions were confirmed or overturned, and appends a dated review section to the prediction file. It also adds generalized observations to rubric-memo.md. The original prediction text is never edited.
- Collects views, likes, comments, shares and favourites, and calculates derived ratios such as likes-to-views
- Groups top comments into categories and reports their proportions
- Marks each earlier prediction as confirmed or overturned, citing specific numbers
- Appends only a review section (## 复盘) to the prediction file
- Writes generalized rules to rubric-memo.md and keeps sample names and comments out of rubric_notes.md
What you need before installing
The skill assumes you already have a prediction file. It will stop if none exists, so it is not a general analytics report tool. The source describes Claude Code or a compatible SKILL.md agent with Bash, Read, Edit, Write, Glob, Grep and Skill tools.
The skill expects the cheat-on-content project layout, which includes prediction files, rubric_notes.md, rubric-memo.md and .cheat-state.json. For automatic data collection (adapter mode) you also need the cloned cheat-on-content repository, Python with httpx for Bilibili, and platform logins saved locally for Douyin, Xiaohongshu, LinkedIn or WeChat Channels. For manual mode you only need to copy metrics and comments from the platform dashboard.
- Claude Code or a compatible SKILL.md agent
- The cheat-on-content project layout with prediction files and state files
- Manual mode: metrics plus at least 5 (ideally 20) top comments from the dashboard
- Adapter mode: the cloned repository's adapters/perf-data folder and platform logins
Install cheat-retro from the ZIP
Download the ZIP for this skill from the page where you found it, then unzip it. The skill is a folder that contains SKILL.md. For a personal install, place the folder in ~/.claude/skills/. To limit it to one project, place it in that project's .claude/skills/ folder instead.
If you use the Claude apps instead of Claude Code, you can upload the skill under Settings → Capabilities → Skills. Either way, the agent will load the skill from its SKILL.md file.
- Personal install: ~/.claude/skills/
- Project install: the project's .claude/skills/ folder
- Claude apps: Settings → Capabilities → Skills, then upload
Set up adapter mode for automatic data collection (optional)
Manual mode works with no extra setup. Adapter mode fetches numbers and comments for you. Adapters live in the cloned repository's adapters/perf-data folder; the installer does not copy them into the skills folder, so you need to locate that folder in your clone.
Bilibili is the only adapter that needs no login. It needs httpx, which you install from its requirements file. Douyin, Xiaohongshu, LinkedIn and WeChat Channels need a saved login first. If an adapter fails for any reason, the skill falls back to manual mode and does not stop the workflow.
- Bilibili: no login step; the adapter does not touch the .auth/ folder
- Douyin, Xiaohongshu and LinkedIn: log in once with the adapter's crawler.py
- WeChat Channels: log in with review.py and scan the QR code; the login stays local to your content project
- YouTube adapter is marked as planned and not yet available in the source
pip install -r requirements.txt
python <adapter>/crawler.py loginTrigger the skill with natural requests
You can invoke the skill with the slash command or with a plain-language request. The source lists these trigger phrases: "复盘 [path]", "retro this", "T+3d 数据来了", "抓数据 [path]" and "把这篇复盘了". The path can point to a prediction file or to a video folder.
You can override defaults on the command line with the window (3, 5 or 7 days) and the source (manual or adapter). If you give no path, the skill takes the earliest entry in pending_retros from .cheat-state.json.
- "retro this" with the prediction file path
- "T+3d 数据来了" when the three-day numbers are ready
- "抓数据 [path]" to fetch data through an adapter
- Omit the path to let the skill pick the oldest pending review
/cheat-retro <prediction-file> — window: 7 — source: adapterWhat the review looks like and how to read it
The skill first checks that the prediction file has a Published at entry and that enough days have passed. It then records the performance section, with derived ratios that single view counts cannot show. Next it lists which predictions were confirmed and which were overturned, quoting the numbers behind each judgement.
The prediction text itself stays as it was. If you already have a review section, the skill asks whether you are adding to it or correcting it, and it refuses to change the prediction section.
- Performance block: views plus ratios such as likes-to-views and comments-to-views
- Comment clusters with example comments and like counts
- Verified and overturned items, each tied to a specific data point
- Proposed rubric observations, which must be traceable to actual data
Safety notes: scripts, network and login data
cheat-retro runs scripts and uses the network. Its allowed tools include Bash, so review what it runs before you grant access. Adapters fetch data from platforms, so they need network access. Login state is stored locally: Douyin, Xiaohongshu and LinkedIn cookies live in .auth folders, and WeChat Channels login stays in your content project.
The source states that the WeChat adapter does not save raw responses, request bodies, full backend URLs, screenshots, page text, commenter nicknames or internal user IDs. Still, keep login folders out of shared repositories.
- Bash access means the agent can run commands in your environment
- Platform logins are stored locally, not in the skill
- Adapter failures fall back to manual mode
- Do not paste private account credentials into the chat
Tips for getting a useful review
Always gather comments. The source treats top comments as the key signal, and it refuses to continue with fewer than 5. Aim for 20 top comments with their like counts. If comments are unavailable, tell the skill why, and it will mark the review as lower value rather than skipping comments.
Choose the window to suit the platform. The default is 3 days, which the source suggests may be shortened for fast short-video platforms and lengthened for long articles. Reviews that are run early are flagged as early_retro.
- Paste at least 5 comments in manual mode, ideally 20
- Use the window setting to match the platform's pace
- Keep prediction files unchanged; add corrections as new review sub-sections
- Watch for repeated bias: the skill proposes a bump review after consistent deviations
What people use it for
- Checking whether a video's predicted reach was too high or too low three days after publishing
- Running the same calibration routine across many articles for a content team
- Spotting that a rubric factor, such as emotional resonance, consistently predicts more or less than assumed
- Recording which comment themes, such as memes or quoted phrases, drove engagement on a Bilibili video
Frequently asked questions
Is cheat-retro free to use?
The source lists the skill under the MIT licence, which is a permissive open-source licence. The source does not describe any pricing, so check the licence text on the page before relying on it commercially.
What do I need to install cheat-retro?
You need Claude Code or a compatible SKILL.md agent, the cheat-on-content project layout and a prediction file to review. Adapter mode also needs the cloned repository and, for most platforms, saved logins. Manual mode only needs metrics and top comments copied from the dashboard.
Where do I put the cheat-retro skill folder?
Place it in ~/.claude/skills/ for personal use or in a project's .claude/skills/ folder for one project. In the Claude apps you can instead upload it under Settings → Capabilities → Skills.
Does cheat-retro work without logging in to platforms?
Yes, through manual mode, where you paste the numbers and comments yourself. Among the adapters, only Bilibili is described as login-free. Douyin, Xiaohongshu, LinkedIn and WeChat Channels need a saved local login before the adapter can fetch data.
What if the skill says there is no prediction file?
The skill assumes a prior prediction exists and stops without one. Create the prediction file first, then run the review. This is not a tool for one-off analytics with no prediction.
What should I do if an adapter fails?
The skill falls back to manual mode and tells you the reason, for example expired cookies or a changed interface. Paste the metrics and at least 5 top comments to continue. If the adapter cannot fetch comments, the review is recorded as lower value.
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
cheat-retro suits creators and content teams who write predictions and want a repeatable way to check them against real results. It is a poor fit for one-off analytics or for anyone who cannot share comments, because comments are central to the review. If you want automatic data pulls without managing platform logins, use the Bilibili path or choose a different tool.
Download the skill