Guide · Agent skill

How to use the cheat-recommend agent skill: pick your next topic

Updated 2026-11-02 · written with AI from the project's own documentation and checked against our catalogue data

cheat-recommend is an agent skill that reads a candidates.md file of topic ideas you have already scored, sorts them by a composite score, and suggests the top few. Each suggestion comes with its scores, a one-sentence reason and comparable past works. It is aimed at creators who keep a backlog of video or post ideas and content teams using the cheat-on-content workflow. After this guide you will know what the skill does, how to install it, how to ask for recommendations and what it needs to work.

Cheat Recommend (Topic Picker)XBuilderLAB/cheat-on-content · MIT Download the skill

Quick start

1Download the ZIP. Get the cheat-recommend skill ZIP from this page and unzip it so that SKILL.md is inside a cheat-recommend folder.
2Place the skill folder. Move the folder to ~/.claude/skills/ for personal use, or to .claude/skills/ in your project. You can also upload the ZIP under Settings → Capabilities → Skills in the Claude apps.
3Prepare candidates.md. Make sure candidates.md sits in your project root with scored entries. If it is missing, the skill will guide you through creating one.
4Ask for a recommendation. Say something like 推荐选题 or recommend topics in a chat opened in your project.
5Apply a filter if needed. Add a request such as top 3 or safe-only if you want fewer picks or no risky topics.
6Review scores and anchors. Check each pick's dimension scores, rationale and anchor before you commit. Challenge any score that looks wrong.
7Make the stable pick. Choose the stable recommendation to produce now, and optionally the experimental one for a test.

An agent skill is a folder that contains a SKILL.md file. The file gives an AI agent such as Claude a set of instructions for one task. When your request matches the skill's trigger words, the agent follows those instructions.

cheat-recommend teaches the agent to read candidates.md, rank the entries by composite score and return a short list. It also skips topics you have already published and topics you marked as skipped, and it adjusts its advice to how much unpublished content you have queued.

Reads candidates.md from your project root and sorts entries by composite score, highest first.
Shows the scores per dimension, a one-sentence rationale and one or two comparable published works for each pick.
Excludes topics that already appear in predictions/*.md and entries marked skip.
Ignores topics that have no score, so it only recommends what has been evaluated.
Flags risky topics and can filter to safe, risky, tier1 or all topics.
If no candidate list exists, it explains how to create one instead of returning an error.

Download the skill ZIP from this page and unzip it. The folder should contain SKILL.md. Then choose one of the two locations below.

For a personal install that works in every project, unzip it into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder.

In the Claude apps you can instead upload the ZIP under Settings → Capabilities → Skills.

Personal: ~/.claude/skills/cheat-recommend/
Project only: .claude/skills/cheat-recommend/ inside your project
Claude apps: Settings → Capabilities → Skills, then upload the ZIP

~/.claude/skills/cheat-recommend/SKILL.md

Check what you need before you start

The skill needs an agent that can read files and search them. The source lists file read, glob and grep access as requirements. Claude or another SKILL.md-compatible agent with those tools is enough.

The main input is a candidates.md file in your project root. Each candidate should be an entry with a title, a tier and a composite score. The skill also reads predictions/*.md and .cheat-state.json from the wider cheat-on-content project, which it uses to remove duplicates and to find comparable past works. Without those files it can still rank candidates, but it cannot exclude published topics or show anchors.

Agent with file read, glob and grep access
candidates.md in the project root, with scored entries
predictions/*.md and .cheat-state.json from cheat-on-content for deduplication and anchors
No scripts are run and no network access is used

Ask for recommendations in natural language

You do not need a command to start. The skill triggers on phrases such as 推荐选题, next topic, 下一篇做什么, recommend topics and 挑一个选题. Any of these in a chat with your project open should bring up the skill.

The source also documents an optional argument format for controlling the result: the number of topics (top N, default 5) and a filter (tier1, all, safe or risky). The format is written as /cheat-recommend — top: 3 — filter: safe.

"推荐选题" or "recommend topics" for a default top 5 list
"下一篇做什么" or "next topic" when you are deciding what to film next
"挑一个选题" when you want a single pick
Filter options: tier1 keeps only tier1, safe drops risky topics, risky shows only risky topics, all keeps everything

/cheat-recommend — top: 3 — filter: safe

Read the output and check the scores

The first pick is a stable choice, meant to be made now. When there are at least two recommendations, the second pick is experimental and is chosen to add useful information, not necessarily to score highest. The remaining slots are filled by composite score and labelled as alternatives.

Every pick shows its dimension scores, a rough predicted length bucket, a rationale, an anchor and any risk. The skill's own guidance is to challenge these scores rather than trust them. If an anchor looks wrong for your channel, that is a signal the composite may be off.

Stable pick: highest composite, safe, not in a recently used category
Experimental pick: tests a hypothesis or a dimension combination you have not tried much
Dimension scores let you see which factor drove the rank
Anchors are published works with similar composite scores, used as a reality check
Buffer colour can change the advice: red asks for one stable pick, blue pauses recommendations until you publish or review queued work

Troubleshooting and setup when candidates.md is missing

If candidates.md does not exist or has fewer than one entry, the skill does not fail. It returns guidance on how to build a candidate pool. The source lists four routes: running the cheat-seed command for a first brainstorm, running cheat-trends to pull scored candidates from sources, adding titles by hand to candidates.md, or importing from Notion or RSS with cheat-init in add-pool mode.

If the skill rejects a file, the cause is usually a format change. The source says the skill asks you about the schema rather than silently ignoring entries it does not recognise. Keep the format used by the candidate schema: an H3 heading per entry with a tier tag, an id and a composite line.

No candidates.md: follow the setup guidance, then ask again
Entries without a composite score are not recommended
Hand-edited files may be questioned by the skill rather than silently skipped

### [tier1] Title of the topic
- **id**: a3f2c1d4e5b6
- **composite (v2)**: 8.47 — ER=4 HP=4 QL=5 NA=3 AB=5 SR=3 SAT=3

What people use it for

Choosing the next video from a backlog of scored ideas when you do not want to decide by gut feeling.
Giving a content team one shared, explained ranking of candidate topics.
Checking whether a risky topic is worth the risk by asking for a risky-only list.
Keeping a posting schedule steady by asking for a single stable pick when your queue is short.

Frequently asked questions

Is cheat-recommend free to use?

The source gives the licence as MIT, and the page lists it as MIT licensed. It does not describe pricing, so check the repository for any terms beyond the licence.

What do I need to install cheat-recommend?

You need an agent that reads SKILL.md files, such as Claude, with file read, glob and grep access. Place the unzipped folder in ~/.claude/skills/ or .claude/skills/, or upload the ZIP in the Claude apps under Settings → Capabilities → Skills.

Why does cheat-recommend say no candidates were found?

It needs a candidates.md file in the project root that has at least one entry, and it only ranks entries with a composite score. If the file is missing or empty, the skill gives setup steps rather than an error.

Does cheat-recommend run scripts or use the internet?

No. The source lists Runs scripts as no and Uses network as no. The skill reads files in your project.

Are the scores accurate?

The scores are heuristic estimates. The skill itself says they should be challenged, not trusted blindly, so compare them against the anchors it shows.

How does it compare to just sorting my list by hand?

The skill adds deduplication against published work, skips skipped and unscored entries, shows dimension scores and anchors, and adjusts advice to your queue. A manual sort does none of this, but it also does not need an agent.

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-recommend if you already keep a scored list of topic ideas and want a consistent, explained ranking with checks against what you have published. Skip it if you have no scored candidates, since it will only send you to setup steps, or if you need guaranteed audience results, since its scores are estimates.

Download the skill