YouTube Research Skills for AI Agents

Practical AI agent skills for YouTube research, powered by the TubeLab MCP.

These skills help agents find niches that are working, bend a proven format into a new market, explain why a video broke out, write trend reports, generate ideas and titles for a channel, analyze a channel, reverse engineer how a channel grew, research thumbnails, map competitors, mine comments and transcripts, audit monetization, find sponsors, build sponsor prospect lists for brands, and watch a niche over time.

Together they work like a full YouTube agency with six roles: a niche scout, a validator, a trend analyst, a script analyst, a packaging team and a monetization team. It runs on your machine or in your chat app, through your own agent and your own TubeLab account.

This is not endpoint routing. Each skill is a short research workflow: the judgment an expert strategist brings, the traps to avoid, and the report to produce. Every number comes from a tool response, every finding carries a linked channel and title, a view count, an outlier ratio and a date.

Install

npx skills add TubeLabHQ/youtube-research-skills

Works with Claude Code, Cursor, OpenAI Codex, GitHub Copilot, Gemini CLI, Windsurf, VS Code, Grok Build, and other agents that support the Agent Skills spec.

Grok Build reads skills from .grok/skills/ in a repo or ~/.grok/skills/, so a copy of the skills/ folder is enough:

git clone https://github.com/TubeLabHQ/youtube-research-skills && cp -r youtube-research-skills/skills/* ~/.grok/skills/

Claude Code can also install it as a plugin, which registers the MCP at the same time:

claude plugin marketplace add TubeLabHQ/youtube-research-skills
claude plugin install youtube-research@tubelab

Setup

Connect the TubeLab MCP once in your client: https://public-api.tubelab.net/mcp

Claude Code: the plugin does it, or claude mcp add --transport http tubelab https://public-api.tubelab.net/mcp. The first call opens OAuth.
Claude.ai and Claude Desktop: the TubeLab connector, OAuth.
ChatGPT and Codex: the TubeLab app, OAuth. Codex supports ChatGPT apps, so no API key is needed there.
Grok Bot: ask a Bot to add the MCP server at the URL above (a Bot can add an MCP server that is not in its catalog by URL) and complete the authentication. Then hand the Bot this repo URL, or paste a SKILL.md, and ask it to save each one as a skill, starting with tubelab-mcp. Skills are shared by all your Bots; if a Bot cannot find one, open Connect apps and enable it under Private skills.
Grok Build (terminal): skills from the folders above; add the MCP from the /mcps modal.
Cursor and other IDE agents: a TubeLab API key as Authorization: Bearer <key> (Pro plan).
Claude.ai and Claude Desktop also take skills directly: zip a skill folder (cd skills && zip -r niche-finder.zip niche-finder/), upload it under Customize > Skills, turn it on, and Claude loads it when a request matches. Upload tubelab-mcp too; the workflow skills name it.
ChatGPT desktop app and Codex: skills live in the Skills sidebar and in .agents/skills/ of a repo; Codex also installs them with $skill-installer.
Any chat client that cannot install skills: paste skills/tubelab-mcp/SKILL.md and then the skill you want into the chat. Every skill is one file, so this works as is.

Setup guides per client: tubelab.net/docs/api/mcp-server. The MCP is included from the Starter plan. Searches cost credits (channels 5, outliers 2, trending formats 1); channel, video, transcript and comment reads are free.

Available Skills

🔭 Niche scout

Skill	Use it when you want to...	Output
`niche-finder`	Find breakout channels and cash cows with two preset searches, and turn them into new niche ideas	Ten ranked niche ideas, each with proof and a way to earn beyond AdSense
`niche-bending`	Bend the proven format of one channel, or of channels that already earn, into a new market or language	Four to six ranked bends, each shown as the source niche and the bend with their thumbnails, with proof, who already does it, and first titles
`niche-watchlist`	Watch a niche and get the delta since the last run, on a schedule	Dated digest of new and moving channels, each with a short analysis: worth studying or not, working topics, title formulas, best recent video. Config and state files

⚖️ Validator

Skill	Use it when you want to...	Output
`niche-validator`	Decide go or no-go on one niche, or compare a shortlist of up to three, with the Niche Analyzer scores and the agent's own read next to them; a new idea is scored on the two niches it draws on	Market Size, Unsaturated and Profitable scores, formula and own read, verdict, channels breaking in, formats, gaps, bending ideas, entry plan; a comparison table first for a shortlist

📈 Trend analyst

Skill	Use it when you want to...	Output
`trend-report`	See what is working in a niche or for a channel, and why	Up to five winning videos with verbatim hooks and why they worked, patterns, ideas

🎬 Script analyst

Skill	Use it when you want to...	Output
`script-breakdown`	Break down a video's structure, hook and beats from its transcript. Given a niche or a channel, it picks the top videos itself	Hook verbatim, typed beats with linked timestamps, one-line layout, optional skeleton

🎨 Packaging team

Skill	Use it when you want to...	Output
`title-formulas`	Get titles you can publish, built on formulas from the trending index, your niche and neighbouring niches. Pairs with `thumbnail-concepts`: titles first	About eight titles with proof, plus a table of their formulas with each one's source, proven or untested
`thumbnail-concepts`	Get thumbnail concepts that join your video's new angle with the problem your viewers know. Pairs with `title-formulas`: the thumbnail for the chosen title	50 or more thumbnails viewed, the patterns in the ones worth studying, three concepts that combine them, images or prompts

💰 Monetization team

Skill	Use it when you want to...	Output
`monetization-research`	Find monetization gaps and what to sell for your channel (the creator's side), from how peers and same-audience channels earn	Table of monetization method gaps, best first, with linked peer offers; the brands that sponsor channels like yours; three products to test
`sponsorship-prospecting`	Find YouTube channels to sponsor for your product, matched to your buyer (the brand's side)	10 to 15 scored prospects with audience, current sponsors and contact path; the best fits with an outreach angle and a cost band; a CSV

🧭 Channel strategist

Skill	Use it when you want to...	Output
`channel-analysis`	Analyze a channel: positioning, what works, what is holding it back	One-line diagnosis, what works, what holds it back, next 3 tests
`reverse-engineer-channel`	Reverse engineer how a channel you admire grew, from its whole video list: turning points, formats, thumbnails and hooks	Growth timeline, turning points, formats, the best ten videos against the worst ten, hooks word for word, how it earns, five moves to copy, each with the skill and prompt to run next, and the moves to skip
`competitor-research`	Compare a channel against 12 to 15 real neighbours: peers, leaders and fast risers	How the set was built, comparison table, gaps, positioning

💬 Audience researcher

Skill	Use it when you want to...	Output
`comment-mining`	Mine top comments for questions, objections, requests and audience language	Themes with verbatim quotes and likes, top comments, the audience's words, video ideas

💡 Ideas

Skill	Use it when you want to...	Output
`video-ideation`	Get ranked video ideas for one channel from three sources: its own winning formats, its niche and related niches, plus same-audience channels on request	About ten ideas, each with a twist and linked proof, ranked by potential and fit

🔌 Data layer

Skill	Use it when you want to...	Output
`tubelab-mcp`	Drive the TubeLab tools well: costs, craft rules, quirks found in testing, and where the full API reference lives	Rules the other skills rely on; no artifact

Every skill is one SKILL.md. Real outputs for every skill are in examples/. Ordered chains for common jobs are in playbooks/, including youtube-agency.md, which runs all six roles from a blank page to a first video and a monetization plan.

The Agency (Claude Code agents)

The plugin also ships six agents, one per role. Each one is a thin wrapper: it loads the skills for its role and returns the finished report.

Agent	What it does	Skills it loads
🔭 `niche-scout`	Finds breakout channels and cash cows and turns them into ten niche ideas, bends a proven format into new markets, or watches a niche over time	`niche-finder`, `niche-bending`, `niche-watchlist`
⚖️ `validator`	Scores a niche, or a shortlist of up to three, and says go or no-go	`niche-validator`
📈 `trend-analyst`	Pulls what is winning right now, with the hooks word for word	`trend-report`
🎬 `script-analyst`	Breaks the top videos down beat by beat	`script-breakdown`
🎨 `packaging-team`	Writes titles, then thumbnails for the chosen title, from what is already getting clicked	`title-formulas`, `thumbnail-concepts`
💰 `monetization-team`	Shows what channels like yours sell and who sponsors them, or which channels a brand should sponsor	`monetization-research`, `sponsorship-prospecting`

In Claude Code, call one by name, for example @agent-youtube-research:niche-scout find me faceless niche ideas. Other clients do not load these agent files. They use the same skills directly, and the reports are the same.

Example Prompts

Find me unsaturated faceless niche ideas in English with real money in them, ranked.

Bend @channel's best format into three new markets. Football first.

Run the whole agency for me: find niches, validate the top three, then give me the trend report, script breakdown, title, thumbnail and monetization plan for the winner.

Trend report for @mychannel this month. Hooks word for word, and ideas I can make.

Analyze @mychannel. Be direct. Why did the last six videos underperform?

Reverse engineer how @channel grew. What can I copy, and what should I skip?

Thumbnail concepts for "How I Would Build a $10M Service Business (If I Had to Start Over)". Here is the script.

Who could sponsor my channel? Find the brands already paying channels my size with my audience.

Who are my real competitors and how do I compare? youtube.com/@handle

We sell an invoicing app to US freelancers. Which YouTube channels should we sponsor?

How the Skills Work Together

tubelab-mcp
        │
        ▼
YouTube research agency
 ├─ 🔭 niche scout          niche-finder · niche-bending · niche-watchlist
 ├─ ⚖️ validator            niche-validator
 ├─ 📈 trend analyst        trend-report
 ├─ 🎬 script analyst       script-breakdown
 ├─ 🎨 packaging team       title-formulas · thumbnail-concepts
 ├─ 💰 monetization team    monetization-research · sponsorship-prospecting
 ├─ 🧭 channel strategist   channel-analysis · reverse-engineer-channel · competitor-research
 ├─ 💬 audience researcher  comment-mining
 └─ 💡 ideas                video-ideation

The workflow skills lean on tubelab-mcp for tool costs, craft rules and quirks. The full parameter and REST reference lives at https://tubelab.net/docs/skill.md. Each workflow produces an artifact a creator can act on or a studio can hand to a client, not raw JSON.

Design Principles

Workflow-first, tool-second. Users ask for an outcome, not a search.
Outlier over raw views. A 50K video on a 5K-average channel says more than 2M on a giant. Rank by ratio, show both.
Recency first. Every search sets a window. Widen one step at a time, and say so.
Content Authenticity is never off. General and above for niches, outliers, ideas and titles; High for anything thumbnail-related. Never relaxed to widen a thin result. The one exception is niche-watchlist, whose saved search leaves it out.
Cited outputs. Every finding carries channel · title verbatim · views · ratio · date. Everything cited is linked with a URL from tool data: videos, channels, moments, brands and products. A sponsor that was only heard links to the video moment. Nothing invented.
Estimates are estimates. Revenue, RPM and audience are model outputs, shown as labelled bands.
Credit-aware. Searches cost credits, reads are free: go deep with free reads, never loop a paid search.
Public data only. Everything comes from public YouTube pages via TubeLab.
Plain language. Skills and reports use short, full sentences that anyone can follow on first read.

youtube-research-skills