How to Build a 3-Agent Claude TikTok System That Learns What Goes Viral
Most people use Claude for TikTok in almost the same way.
They ask for content ideas.
Claude generates 20 or 30 topics.
They choose one.
Claude writes a script.
They publish it.
The next day, they come back and ask Claude for another batch of ideas.
There is a major weakness in this approach:
Claude usually has no idea what happened after you published.
Maybe one hook doubled your average watch time.
Maybe your personal stories generated three times more shares than your tutorials.
Maybe viewers ignored your “5 Tips” videos but saved your step-by-step demonstrations.
Maybe a video with 40,000 views brought you far more interested followers than one that reached 500,000.
If those lessons never make their way back into your AI workflow, every new content session starts almost from scratch.
There is a better way to use Claude.
You can create a 3-agent Claude TikTok system where:
Agent 1 researches content opportunities and audience demand.
Agent 2 turns the strongest ideas into TikTok scripts using your real voice, experiences, stories, and opinions.
Agent 3 analyzes what happened after publication and feeds those lessons back into the next round of research and writing.
Instead of repeatedly asking AI to guess what might work, you create a feedback loop:
Research → Score Ideas → Write → Publish → Measure → Learn → Research Again
The goal is not to create an AI machine that magically predicts every viral TikTok.
Nobody can reliably promise that.
The real goal is much more useful:
Build an AI content system that gets better informed about what works with your particular audience each time you publish.
This tutorial shows you how to build that system from scratch in 10 steps.
By the end, you'll have:
- A Claude Cowork workspace
- A personal voice database
- A story and opinion library
- A TikTok research process
- A content scoring method
- A scriptwriting agent
- A TikTok performance tracker
- A performance-analysis agent
- A learning database
- A weekly feedback loop
You can build the first version without coding.
How Does a 3-Agent Claude TikTok System Work?
The system divides TikTok content creation into three jobs.
Agent 1: Content Research Agent
Agent 1 studies:
- TikTok searches
- content gaps
- current trends
- successful content formats
- recurring audience questions
- your previous results
Its job is to identify what might be worth talking about.
Agent 2: TikTok Scriptwriting Agent
Agent 2 takes the strongest ideas and combines them with:
- your voice
- your experiences
- your stories
- your opinions
- your audience profile
- your previous winning patterns
Its job is to determine how you should communicate the idea.
Agent 3: Performance Learning Agent
Agent 3 studies:
- views
- watch time
- retention
- comments
- shares
- saves
- profile activity
- follower growth
- business results when available
Its job is to determine:
What did your audience teach you?
Those lessons go back to Agents 1 and 2.
That closes the loop.
What You Need Before Starting
The basic version requires very little.
You need:
- A paid Claude plan with access to Claude Cowork
- A TikTok account
- A niche or subject you want to discuss
- Some personal knowledge or experience
- A place to store your content files
- TikTok performance data once you begin publishing
Claude Cowork is useful for this workflow since it can work with project files, instructions, memory, scheduled tasks, plugins, and sub-agents.
You do not need to start with complicated autonomous agents.
For the beginner version, we will create three clearly defined AI roles working from the same knowledge base.
Once the basic system works, you can move into scheduled tasks, plugins, and true Claude sub-agents.
Step 1: Train Claude on Your Voice, Stories, and Opinions
Before Claude can write good TikTok scripts for you, it needs something most AI content systems never receive:
A detailed picture of you.
Not this:
I'm a marketing consultant. Write casually.
Claude needs much more.
It needs:
- your stories
- your opinions
- your vocabulary
- your experiences
- your mistakes
- your humor
- your disagreements
- your speaking habits
- your preferences
- your boundaries
- the phrases you hate
- the things you would never say
One useful approach comes from AI creator Ruben Hassid, who has discussed using a 100-question interview to capture someone's taste, thinking, preferences, and communication style.
The key is not the number 100 itself.
The value comes from asking enough thoughtful questions that Claude begins to see patterns that would never appear in a short “brand voice” description.
Run a 100-Question Personal Interview
Open a new Claude conversation and use this prompt:
I want you to interview me deeply so you can learn how I think, speak, tell stories, explain ideas, and communicate.
Ask me 100 questions, one at a time.
The purpose is to build a detailed personal knowledge and voice profile that another AI could later use to help me create short-form social media content that genuinely sounds like me.
Cover these areas:
1. My background and experiences
2. Stories from my life and work
3. Strong opinions
4. Beliefs about my industry
5. Ideas I disagree with
6. Mistakes I have made
7. Lessons I have learned
8. How I naturally explain things
9. Words and expressions I commonly use
10. Words and expressions I dislike
11. Humor and personality
12. How I tell stories
13. How I open conversations
14. How I make an argument
15. Topics I care deeply about
16. Topics I do not want associated with me
17. Things I would never say
18. AI-writing patterns that sound fake to me
19. My ideal audience
20. What I want people to feel after hearing me
Do not ask all 100 questions at once.
Ask one question, wait for my answer, and then ask the next.
If my answer is vague, ask a useful follow-up before moving forward.
If I mention an interesting story, opinion, contradiction, failure, insight, or experience, explore it.
Do not try to improve my answers during the interview. Preserve the way I actually communicate.
At the end, organize what you learned into a master profile.
Speak Your Answers When Possible
Voice input can work very well for this exercise.
When people type, they often edit themselves.
They shorten stories.
They replace ordinary phrases with more formal language.
They clean up unfinished thoughts.
That may create better written prose, but it can remove the exact patterns Claude needs to learn how you naturally communicate.
Talk the way you normally talk.
If a question reminds you of a story, tell the story.
If you disagree with the premise, say so.
If you are uncertain, say that too.
Claude does not need a polished interview.
It needs an accurate one.
Give Claude Your Real Opinions
Generic AI content often feels generic for one simple reason:
The AI has nothing distinctive to work with.
Suppose ten creators can explain the same marketing tactic.
Why should someone listen to you?
Maybe you tried the tactic and discovered an unexpected problem.
Maybe you disagree with the standard advice.
Maybe you lost money doing what everyone recommended.
Maybe you have used the strategy for 15 years and noticed something newcomers miss.
Maybe your experience led you to a very different conclusion.
Those are the ingredients that can make content worth watching.
Step 2: Turn Your Interview Into a Personal Content Database
Once the interview is complete, do not leave everything trapped inside one giant conversation.
Turn the interview into reusable files.
Ask Claude:
Using everything from my interview, create the following Markdown files.
1. master-interview.md
Preserve the important substance of my answers, including stories, opinions, experiences, examples, contradictions, preferences, and unusual observations.
2. voice-profile.md
Describe how I naturally communicate: sentence patterns, tone, vocabulary, humor, storytelling habits, preferred explanations, pacing, and conversational characteristics.
3. stories.md
Create a story bank containing every useful personal, professional, emotional, funny, surprising, embarrassing, educational, or meaningful story from the interview.
4. beliefs-opinions.md
List my important beliefs, opinions, disagreements, contrarian views, lessons, and observations.
5. anti-ai-rules.md
List writing patterns, phrases, structures, clichés, transitions, hooks, and habits that would make content sound artificial or unlike me.
6. hard-boundaries.md
List things I do not want AI to invent, exaggerate, claim, discuss, or associate with me.
Do not invent information.
Preserve uncertainty when I expressed uncertainty.
Keep concrete examples wherever possible.
You now have something more useful than a single “write in my style” prompt.
You have a personal content database.
Create an Audience Profile
Claude needs to know who you're trying to reach too.
Create:
audience-profile.md
Use this prompt:
Help me build a detailed audience profile for my TikTok content.
Based on what you already know about me and my niche, interview me as needed and document:
Who my ideal viewer is
What they want
What frustrates them
What they have already tried
What they misunderstand
What they fear
What they are curious about
What might make them save a video
What might make them share a video
What might make them comment
What might make them follow
What questions they may be searching for on TikTok
What outcome they ultimately want
Do not invent certainty when we do not have enough information.
Create Content Pillars
Your account probably should not talk about everything.
Create three to seven broad content pillars.
An AI marketing creator might use:
- AI tools
- AI workflows
- marketing experiments
- mistakes and lessons
- case studies
- personal business stories
Save them as:
content-pillars.md
Separate Permanent Files From Learning Files
This distinction becomes critical later.
Permanent Files
These describe who you are:
master-interview.md
voice-profile.md
stories.md
beliefs-opinions.md
hard-boundaries.md
audience-profile.md
Do not let an analytics agent rewrite these every week.
Learning Files
These describe what your content experiments are teaching you:
winning-hooks.md
weekly-lessons.md
topic-priorities.md
format-performance.md
audience-observations.md
experiments.md
These are allowed to change.
That prevents one strange viral video from changing your entire identity.
Step 3: Set Up Your Claude Cowork TikTok Workspace
Now create a home for the system.
A practical structure could look like this:
/TIKTOK-CONTENT-SYSTEM
/01-IDENTITY
master-interview.md
voice-profile.md
stories.md
beliefs-opinions.md
anti-ai-rules.md
hard-boundaries.md
/02-AUDIENCE
audience-profile.md
content-pillars.md
audience-problems.md
/03-RESEARCH
trend-research.md
idea-bank.md
winning-hooks.md
/04-SCRIPTS
approved-ideas.md
draft-scripts.md
published-scripts.md
/05-PERFORMANCE
post-performance.csv
weekly-lessons.md
format-performance.md
topic-priorities.md
audience-observations.md
experiments.md
Create a Cowork Project called:
TikTok Content System
Then give Claude project instructions:
This project manages my TikTok content research, ideation, scripting, publishing analysis, and performance learning.
Never invent personal experiences for me.
Use my identity files when writing in my voice.
Treat permanent identity files as protected.
Performance results may update the learning files but should not automatically alter my core beliefs, personal history, hard boundaries, or voice profile.
Our objective is to create useful, original content for my target audience and learn systematically from real publishing results.
Do not assume that the post with the most views is automatically the best post.
Evaluate reach, retention, engagement, audience growth, and business relevance separately.
You now have one shared workspace feeding all three agents.
Step 4: Build Agent 1, Your TikTok Research Agent
Agent 1 has one main responsibility:
Find promising content opportunities.
It should not simply find viral videos and rewrite them.
A stronger formula is:
Audience demand + trend information + content gaps + your experience = original idea
Use TikTok Creator Search Insights
TikTok's Creator Search Insights can surface things such as:
- popular searches
- related topics
- content gaps
- search interest
- related videos
- searches connected to your existing audience
This matters since search demand can reveal problems people actively want solved.
A search such as:
“How to use Claude for marketing”
may have less raw attention than a celebrity trend, yet far more value for someone building an AI marketing audience.
Look for:
- questions your audience asks repeatedly
- high-interest searches
- underserved searches
- weak explanations
- topics where you have a different viewpoint
Use TikTok Creative Center
TikTok Creative Center can help you study current hashtags, trends, related videos, and creative patterns.
Don't stop at:
“This hashtag is trending.”
Ask:
- What opening styles keep appearing?
- How long are the stronger videos?
- Are they stories, tutorials, demonstrations, reactions, or opinions?
- What creates curiosity?
- What makes viewers comment?
- What makes viewers share?
- What seems repetitive?
- What perspective is missing?
Create a Research Card
For each promising opportunity, record:
CONTENT RESEARCH CARD
Topic:
Search phrase:
Trend:
Content gap:
Video URL:
Creator:
Date:
Length:
HOOK
Opening concept:
Hook type:
Curiosity created:
STRUCTURE
Opening:
Setup:
Core value:
Payoff:
CTA:
PSYCHOLOGY
Main emotion:
Problem addressed:
Desired result:
Reason someone might share:
Reason someone might comment:
DIFFERENTIATION
What most creators are saying:
What I believe:
Relevant personal story:
Original angle available:
PRODUCTION
Talking head:
Screen recording:
Demonstration:
Voiceover:
B-roll:
Other:
Agent 1 Prompt
Use this as the core instruction:
You are my Content Research Agent.
Your job is to identify promising TikTok content opportunities for my audience.
Before generating ideas, read:
audience-profile.md
content-pillars.md
voice-profile.md
stories.md
beliefs-opinions.md
weekly-lessons.md
winning-hooks.md
topic-priorities.md
experiments.md
Study current search demand, content gaps, trends, recurring audience questions, successful content structures, and emerging conversations relevant to my niche.
Do not copy another creator's script.
For every opportunity, ask:
1. What is the audience interested in?
2. Why is this interesting now?
3. What problem or desire does it connect to?
4. What is missing from existing content?
5. Do I have a relevant story, experience, opinion, or viewpoint?
6. Can we say something meaningfully different?
7. What would make the idea worth sharing or saving?
Generate 30 candidate content ideas.
For each idea provide:
working title
audience problem
proposed hook
format
unique angle
personal story or opinion available
expected viewer payoff
reason someone might share it
reason someone might follow after watching it
Do not write full scripts yet.
Agent 1 has now given you 30 ideas.
Do not send all 30 directly into production.
Step 5: Score and Select the Best Content Ideas
Generating ideas is cheap.
Producing good content still takes time.
Score the ideas before writing complete scripts.
Give every idea a 1-to-10 score for:
- audience relevance
- hook potential
- originality
- story potential
- emotional interest
- share potential
- save potential
- comment potential
- follower potential
- business relevance
Use:
Score each of the 30 ideas from 1 to 10 for:
Audience relevance
Hook potential
Originality
Story potential
Emotional interest
Share potential
Save potential
Comment potential
Follower potential
Business relevance
Explain the biggest strength and weakness of each idea.
Then rank all 30.
Select the 10 strongest ideas.
Do not automatically favor broad topics.
Prefer ideas where audience demand intersects with something distinctive I can contribute.
Now you have an editorial filter.
Instead of:
30 ideas → 30 scripts
you have:
30 ideas → 10 serious candidates
That can remove a huge amount of mediocre AI content before it reaches production.
Step 6: Build Agent 2, Your TikTok Scriptwriting Agent
Agent 2 receives your highest-scoring ideas.
Its job is to find the strongest way to communicate each one.
Give Agent 2 access to:
voice-profile.md
stories.md
beliefs-opinions.md
anti-ai-rules.md
hard-boundaries.md
audience-profile.md
winning-hooks.md
approved-ideas.md
Generate Hooks Before Writing the Full Script
Don't immediately ask Claude for one finished video.
Generate multiple hooks first.
Suppose your topic is:
Using three Claude agents to improve TikTok content
Possible openings might include:
“Most people make the same mistake every time they use AI for content.”
“Your AI forgets the most useful thing every time you post.”
“I stopped asking AI for random content ideas.”
“What if Claude studied every TikTok you published?”
The underlying idea is almost identical.
The framing is completely different.
That framing may determine whether someone keeps watching.
Agent 2 Prompt
You are my TikTok Scriptwriting Agent.
Read my identity, audience, story, voice, writing-rule, and performance-learning files before writing.
You will receive an approved content idea.
Before writing the script:
1. Generate 15 substantially different hook concepts.
2. Score each hook for curiosity, clarity, relevance, originality, and fit with my voice.
3. Select the three strongest.
4. Choose the best content format for this idea.
Possible formats include:
personal story
contrarian opinion
tutorial
case study
mistake and lesson
reaction
myth correction
before-and-after
experiment
prediction
screen demonstration
Then write three script variations using the strongest hooks.
The scripts should sound natural when spoken aloud.
Use my real stories and opinions when relevant.
Never invent experiences, income, results, clients, credentials, statistics, or personal claims.
Avoid every pattern listed in anti-ai-rules.md.
Do not force a call to action into every video.
Create a clear viewer payoff.
After writing, run this quality check:
Would I actually say this?
Is any part generic?
Is there a real viewpoint?
Is there something distinctive here?
Can anything be shortened?
Does the opening create immediate interest?
Does the payoff justify the opening?
Does it sound spoken rather than written?
Give every final script a unique script ID.
Assign identifiers such as:
TT-2026-001
TT-2026-002
TT-2026-003
Those IDs become useful once you start analyzing performance.
Step 7: Publish and Track Every TikTok
This is where many AI content workflows fall apart.
People create.
They publish.
Then they rely on memory.
A month later, they say things like:
“I think stories usually perform better.”
That isn't enough.
Create a tracking sheet.
Include fields such as:
Script ID
Publish date
Topic
Content pillar
Hook type
Format
Story / no story
Video length
CTA type
Production style
Views
Likes
Comments
Shares
Saves
Average watch time
Watched full video %
Profile activity
Follower growth
Business result
Notes
Use Consistent Measurement Windows
Don't compare a TikTok that has been live for two hours with one that has been live for ten days.
Choose standard checkpoints.
A practical setup is:
24 Hours
Early performance.
72 Hours
Early comparison.
7 Days
Primary comparison window.
30 Days
Useful for long-tail performance when you want it.
The exact window matters less than consistency.
Track More Than Views
Views are useful.
They are not the whole story.
A video can attract huge reach and almost no meaningful action.
Another can receive modest reach but create:
- more shares
- more saves
- better retention
- more profile visits
- more followers
- more leads
- more customers
That is why Agent 3 needs several kinds of data.
Step 8: Build Agent 3, Your Performance Learning Agent
Agent 3 is what turns this from a script-generation workflow into a learning system.
Suppose:
Video A
500,000 views
5,000 shares
Video B
50,000 views
1,500 shares
Video A has far more total shares.
Yet the share rates tell another story.
Video A:
1% share rate
Video B:
3% share rate
Video B may contain a creative pattern worth testing again.
Calculate Useful Ratios
Have Agent 3 calculate:
Engagement Rate
(likes + comments + shares + saves) / views × 100
Share Rate
shares / views × 100
Save Rate
saves / views × 100
Comment Rate
comments / views × 100
Average Retention
When video duration and average watch time are available:
average watch time / video duration × 100
A 25-second average watch time on a 30-second video tells a very different story from 25 seconds on a two-minute video.
Create Different Types of Winners
Don't create one category called:
Best Post
Create several.
Reach Winner
Generated unusually strong views.
Engagement Winner
Generated unusually strong shares, comments, or saves.
Retention Winner
Held viewer attention especially well.
Growth Winner
Created meaningful profile activity or audience growth.
Business Winner
Generated clicks, leads, subscribers, customers, or another business outcome you can track.
Now you're asking a better question than:
“Which TikTok went viral?”
You're asking:
“What job did this content perform?”
Agent 3 Prompt
You are my TikTok Performance Learning Agent.
Your job is to analyze my actual publishing results and identify repeatable creative lessons.
Read:
post-performance.csv
published-scripts.md
weekly-lessons.md
winning-hooks.md
format-performance.md
topic-priorities.md
audience-observations.md
experiments.md
Analyze posts using comparable measurement windows.
Calculate useful ratios including:
engagement rate
share rate
save rate
comment rate
average retention when possible
Do not judge posts using views alone.
Classify meaningful winners as:
Reach Winner
Engagement Winner
Retention Winner
Growth Winner
Business Winner
Study performance by:
topic
hook type
format
video length
storytelling use
CTA
production style
content pillar
Look for repeated patterns.
Separate strong evidence from weak evidence.
Never treat one unusual post as proof of a permanent rule.
For every conclusion, state:
1. What happened
2. The evidence
3. Your confidence level
4. Whether we should adopt a rule or run another experiment
Generate:
weekly performance summary
winning patterns
weak patterns
promising hypotheses
recommended experiments
recommendations for Agent 1
recommendations for Agent 2
Now Agent 3 can produce observations such as:
Personal stories generated roughly twice the share rate of generic tutorials this week.
Or:
Videos under 40 seconds produced stronger average retention across six comparable posts.
Or:
Contrarian openings generated more comments, but tutorial openings generated more saves.
Or:
AI-tool content generated reach, but business case studies produced more profile activity.
Those lessons can influence what you create next.
Step 9: Feed TikTok Performance Data Back Into Claude
Now close the loop.
Agent 3 updates your learning files.
For example:
winning-hooks.md
CURRENT HIGH-PERFORMING PATTERNS
Personal confession opening
Evidence: 4 videos
Strongest metric: retention
Contrarian statement opening
Evidence: 5 videos
Strongest metric: comments
Direct problem opening
Evidence: 7 videos
Strongest metric: saves
topic-priorities.md
HIGH PRIORITY
AI workflows
Strong search interest
Above-average saves
Personal marketing failures
High shares
Strong comments
LOWER PRIORITY
Generic productivity tips
Adequate views
Low engagement
Little differentiation
Agent 1 reads these files before its next research session.
Agent 2 reads the winning-hook and format lessons before writing.
Now the feedback loop is complete.
Protect Your Identity Files
Suppose you publish a silly joke and it unexpectedly gets two million views.
Agent 3 should not decide:
“Comedy is now your personality.”
One viral outlier is data.
It is not your identity.
Your permanent files remain protected.
Your learning files evolve.
Create experiments.md
Agent 3 will sometimes notice an interesting pattern without having enough evidence to trust it.
Instead of turning it into a rule, create an experiment.
EXPERIMENT 014
Hypothesis:
Personal-story openings increase retention.
Control:
Direct tutorial opening
Variation:
Personal failure story opening
Keep similar:
Topic
Length
CTA
Production quality
Primary metric:
Average retention
Secondary metric:
Share rate
Test:
Three videos of each type
Status:
Open
Now Claude is helping you decide what to test next.
That is far more useful than simply describing what happened last week.
Step 10: Run the 3-Agent TikTok System Every Week
You now have the complete system.
Turn it into a routine.
Monday: Research
Agent 1 studies:
- TikTok searches
- content gaps
- current trends
- audience questions
- last week's lessons
- open experiments
It generates 30 candidate ideas.
Tuesday: Score and Write
Score the ideas.
Select your strongest candidates.
Agent 2 generates multiple hooks and scripts.
You review them.
Edit anything that does not sound right.
Wednesday: Record
Record several videos.
Batch production can save time if it fits your style.
Keep the setup manageable.
You are trying to gather useful publishing experience, not make every 30-second TikTok into a movie production.
Thursday Through Saturday: Publish
Publish according to your schedule.
Record:
- Script ID
- Hook
- Topic
- Format
- Length
Then collect performance data at your standard checkpoints.
Sunday: Analyze
Agent 3 reviews the week's results.
It identifies:
- winners
- weak posts
- repeated patterns
- anomalies
- new hypotheses
- experiments worth running
It updates the learning files.
The next Monday, Agent 1 begins with that new knowledge.
Your system becomes:
Research
↓
Generate
↓
Score
↓
Write
↓
Publish
↓
Measure
↓
Learn
↓
Research Again
Can Claude Automate Parts of the System?
Yes.
Claude Cowork supports recurring scheduled tasks, which can make parts of this process easier.
For example, you could create a Monday research task.
Monday Research Task
Research emerging content opportunities in my niche.
Read my audience profile, topic priorities, winning hooks, weekly lessons, and open experiments first.
Generate and score 30 fresh TikTok content opportunities.
Save the report in the research section of this project.
Then create a weekly performance-review task.
Sunday Review Task
Review the latest available TikTok performance data.
Analyze results using the rules in my Performance Learning Agent instructions.
Identify meaningful patterns.
Update weekly lessons.
Recommend new experiments.
Create recommendations for the next research cycle.
Start manually before automating everything.
You want to know whether the process works before making it more elaborate.
Advanced: Turn the Three Roles Into Claude Sub-Agents
The beginner system works through Cowork instructions and shared files.
Claude can go further.
Its plugin system can support reusable skills, connectors, and sub-agents.
That means an advanced user could package:
Research Agent
Scriptwriting Agent
Performance Agent
into a more formal setup.
I would treat this as phase two.
A weak research process remains weak when automated.
A poor voice database still produces poor scripts.
Missing analytics still leave Agent 3 guessing.
Build the process first.
Then automate the parts that prove valuable.
Example: From TikTok Trend to Learning Loop
Let's see what the whole system might look like in practice.
Suppose your niche is AI marketing.
Agent 1 Finds an Opportunity
Creator Search Insights shows interest around people searching for ways to create social-media content with Claude.
Agent 1 notices most existing content focuses on prompts.
Your beliefs file contains this opinion:
AI prompts are less interesting than AI systems that learn from results.
Agent 1 creates:
Idea
Why Your AI Content Never Gets Smarter
Proposed Hook
“Your AI forgets the most useful thing every time you post.”
Format
Contrarian tutorial
Unique Angle
Most AI advice focuses on generating content.
This video focuses on the feedback loop after publication.
The idea receives a strong score.
Agent 2 Writes the Script
Agent 2 creates 15 possible hooks.
The winning opening is:
“Your AI forgets the most useful thing every time you post.”
The script explains:
You ask AI for ideas.
AI writes the scripts.
You publish them.
One succeeds.
One fails.
Then tomorrow, you ask the AI for another 20 ideas without telling it what happened.
The video introduces the solution:
Research Agent → Script Agent → Performance Agent
You Publish It
Seven days later:
75,000 views
4,200 likes
380 comments
2,800 shares
1,500 saves
Average retention is much higher than your recent account average.
Agent 3 Analyzes the Result
Agent 3 does not simply say:
“This was your best video.”
It identifies:
- strong share rate
- strong save rate
- above-average retention
- contrarian opening
- clear problem in the opening seconds
- framework-style content
- short introduction
- direct payoff
Agent 3 sees only one example, so it does not create a permanent rule.
Instead, it writes:
HYPOTHESIS
Contrarian problem-first hooks may improve retention for AI workflow content.
Confidence:
Moderate
Next test:
Use the same hook architecture on three different AI topics.
Agent 1 looks for three suitable ideas.
Agent 2 writes the scripts.
You publish them.
Agent 3 compares the results.
Now your AI content strategy is accumulating evidence.
Common Mistakes That Can Ruin the System
Mistake 1: Copying Viral Creators
Agent 1 should research patterns, not steal scripts.
The goal is:
Audience demand + successful structure + your perspective
not:
Viral video + AI rewrite
Your personal knowledge files exist to create differentiation.
Mistake 2: Chasing Views Alone
A video with enormous reach can still bring you almost no meaningful audience growth.
Track the outcome you care about.
That may be:
- reach
- retention
- shares
- saves
- followers
- leads
- subscribers
- sales
Different TikToks can serve different jobs.
Mistake 3: Giving Agent 3 Too Little Data
Five posts can produce interesting clues.
They usually cannot prove universal rules.
Let the dataset grow.
Turn weak evidence into experiments rather than permanent instructions.
Mistake 4: Overreacting to One Viral Video
Outliers happen.
A news event can push a topic.
TikTok may distribute one post unusually widely.
A celebrity reference can create temporary interest.
Treat individual winners as clues.
Look for repeated patterns.
Mistake 5: Making Every TikTok Look the Same
Once Claude identifies a strong format, it may start repeating it.
Keep experimenting.
Try:
- stories
- tutorials
- opinions
- demonstrations
- case studies
- reactions
- experiments
- mistakes
- predictions
Different ideas deserve different formats.
Mistake 6: Letting Claude Invent Personal Stories
Your story database is one of the strongest parts of this system.
It becomes useless if Claude begins making things up.
Never allow Agent 2 to invent:
- clients
- revenue
- failures
- achievements
- personal experiences
- conversations
- credentials
If Claude needs a story and none exists, it should tell you that.
Mistake 7: Removing Human Judgment
Claude may score an idea 97 out of 100.
You may hate it.
Don't make it.
Claude may recommend discussing a personal experience you prefer to keep private.
Don't publish it.
Claude may discover that manipulative hooks attract clicks.
You do not need to use them.
The system should help you make better creative decisions.
It does not need to make every decision for you.
Frequently Asked Questions
Can Claude create TikTok videos automatically?
Claude can help research topics, generate ideas, write scripts, organize files, analyze performance data, and automate parts of a recurring content workflow.
Recording, editing, publishing, and retrieving analytics may still involve human action or separate tools depending on your setup.
Do I need Claude Cowork?
No.
You can reproduce much of the basic workflow with regular Claude chats and organized files.
Cowork makes the system easier to manage once you want persistent project context, recurring tasks, plugins, and more advanced workflows.
What are the three Claude agents?
The three roles are:
Content Research Agent
Finds promising subjects and content opportunities.
TikTok Scriptwriting Agent
Turns selected ideas into hooks and scripts using your voice and knowledge.
Performance Learning Agent
Studies the results and feeds lessons back into future research and writing.
Can Claude predict which TikTok will go viral?
No.
Claude can analyze patterns and help you make better-informed decisions.
TikTok performance remains unpredictable.
This system learns which creative choices have been working more often with your audience.
How many TikToks do I need before Agent 3 becomes useful?
You can begin analyzing immediately.
A few posts can produce hypotheses.
Larger samples can reveal more dependable patterns.
The key is to distinguish between:
“This happened once.”
and:
“This keeps happening.”
Should Agent 3 automatically update my voice profile?
No.
Keep your permanent identity files separate from your adaptive learning files.
Agent 3 can update:
- winning hooks
- format observations
- topic priorities
- audience observations
- experiments
It should not automatically rewrite:
- personal history
- stories
- beliefs
- voice
- boundaries
Can I use this system for YouTube Shorts or Instagram Reels?
Yes.
The same basic structure can work for:
- Instagram Reels
- YouTube Shorts
- Facebook Reels
- LinkedIn video
- other short-form platforms
The platform-specific research methods and analytics will need to change.
Do I need coding skills?
No.
The beginner version can be built using Claude Cowork, text files, prompts, and a simple spreadsheet.
Coding becomes relevant only if you later decide to build more advanced integrations or automation.
What “Learns What Goes Viral” Really Means
The title of this article is:
How to Build a 3-Agent Claude TikTok System That Learns What Goes Viral
That does not mean Claude discovers a secret TikTok formula.
The system is learning something much more realistic:
Which topics, stories, hooks, formats, lengths, creative choices, and ideas have been performing better with your particular audience.
That knowledge accumulates.
After five videos, Agent 3 knows very little.
After 25 videos, patterns may begin appearing.
After 100 videos, your system may contain a detailed history of:
- topics
- hooks
- scripts
- formats
- experiments
- failures
- winners
- audience reactions
- performance patterns
Compare that with someone who has published 100 TikToks yet still begins every AI session with:
“Give me 20 viral TikTok ideas.”
They may have months of publishing experience.
Their AI knows almost nothing about those months.
Your system does.
That is the real advantage.
Final Thoughts: Build the First Version Before You Automate It
Start with the simplest version.
Create your voice files.
Build the three roles.
Research real audience demand.
Score your ideas.
Write the scripts.
Publish several videos.
Track what happens.
Give the results back to Claude.
See what it discovers.
Then repeat.
Once the workflow begins producing useful lessons, you can start experimenting with scheduled research, automated weekly reports, Claude plugins, and sub-agents.
The most valuable part of this system isn't having three AI agents.
It's creating a record of what your audience keeps teaching you and feeding those lessons back into what you create next.
Your first version may be imperfect.
That is expected.
The entire point of the system is that version two knows something version one did not.
Version ten should know much more than version two.
At that point, Claude is no longer just generating more content for you.
It is helping you learn from the content you've already published.
Leave a Reply