Most people ask an AI for content ideas far too early.
If the agent has not watched the videos, compared the creator’s best posts with their normal posts and read the audience comments, it is not doing strategy. It is guessing.
This is the exact setup I use to make an agent research first and generate ideas second.
Before building Lola OS, I ran Naviro, a social media automation and data-analysis platform. The principle is the same here: give the agent real evidence, tell it exactly what counts as useful and make it show its work.
You do not need to understand databases, scoring systems or code to use this workflow.
What you need
- An AI coding agent that can use your terminal
- An Apify account
- A list of creators you genuinely want to learn from
- One clear sentence describing your audience
- A short list of things you can credibly build, teach or explain
I use /goal so the agent keeps working until the research and final output pass a checklist. If your agent does not support /goal, replace the first line with: “Keep working until every acceptance check below passes.”
Step 1: Set up Apify
Create an Apify account, then open Settings → API & Integrations.
Create a token for this project and copy it. A separate token is easier to rotate later if it is ever exposed.
Do not paste the token into your prompt, a public document or GitHub.
Step 2: Store the API key in your terminal
On Mac or Linux, open the terminal where you will run your agent and paste:
export APIFY_API_TOKEN="paste_your_token_here"Check that it is available without printing the secret:
test -n "$APIFY_API_TOKEN" && echo "Apify is connected"On Windows PowerShell, use:
$env:APIFY_API_TOKEN="paste_your_token_here"This stores the key for the current terminal session. If you want it available every time you open the project, ask your agent to create a local .env file and add .env to .gitignore.
Step 3: Give the agent a useful brief
Do not begin with “give me viral content ideas.” First tell the agent who the work is for, whose formats you respect and what you can actually make.
Copy this and replace the brackets:
MY ACCOUNT
@[your handle]
MY AUDIENCE
[Who they are, what they are trying to become and what currently feels difficult.]
CREATORS TO STUDY
@[creator 1]
@[creator 2]
@[creator 3]
@[creator 4]
@[creator 5]
WHAT I CAN CREDIBLY MAKE
[Your skills, projects, professional experience and subjects you can explain.]
DO NOT RECOMMEND
[Topics, tones and formats that feel wrong for you.]
THE OUTCOME I WANT
[For example: three TikTok ideas I can film this week, each with a proven hook structure, a simple script and something useful the viewer can keep.]Here is the brief I would use for myself:
MY ACCOUNT
@lolasasfi
MY AUDIENCE
Women in their twenties and early thirties who work in creative, marketing, product, operations or adjacent careers and want to become confident using AI and technology without becoming programmers.
CREATORS TO STUDY
@morillui
@meshtimes
@ubeboobey
@catmctp
@gazi
WHAT I CAN CREDIBLY MAKE
No-code AI skills, small useful apps, vibe-coded experiments, content-research systems, social-data analysis, creative technology and overlooked women in computing history.
DO NOT RECOMMEND
Generic AI news, coding lessons, founder diary posts, vague productivity advice or content about my private frustration that gives the viewer nothing.
THE OUTCOME I WANT
Three TikTok ideas I can film quickly. Each idea must teach something genuinely useful, open with visible proof and become an article, carousel or downloadable resource.Step 4: Give the agent this /goal prompt
Paste your completed brief above this prompt, then run the whole thing as one task:
/goal Build a research-backed TikTok content plan from real evidence. Do not stop until every acceptance check at the bottom passes.
Use the APIFY_API_TOKEN already stored in the terminal. Use Apify to collect TikTok profile posts, keyword-search results and comments.
RESEARCH
1. Study the creators in my brief and find additional creators with a similar audience, taste and level of technical creativity.
2. Collect at least 20 posts per creator.
3. Use Python—not mental maths—to compare each post with that creator’s normal performance. Do not rank creators only by raw views.
4. Watch at least the five strongest posts from each core creator, plus two ordinary or weak posts for comparison.
5. For every watched post, record:
- the exact first spoken sentence
- the exact opening text on screen
- what appears in the first frame
- when the creator moves from hook to explanation, proof, story and CTA
- the editing and reveal pattern
- the post link and public metrics
6. Collect comments from the relevant posts. Save exact questions, confusions, desires and requests in the audience’s own words.
7. Keep every hook, comment, transcript and screenshot attached to the post it came from. Never use evidence from one post to support another.
8. Save the research so I can later ask questions such as: “Show me every skill tutorial and the exact moment it moves from the hook into proof.”
CREATE THE IDEAS
9. Identify the hook and format patterns that repeatedly appear in strong posts but not in the comparison posts.
10. Generate 20 new ideas that connect:
- one exact audience question
- one subject I can credibly teach or build
- one proven creator format
11. Do not repeat ideas from my brief and do not give me topics. Give me filmable concepts.
12. For every proposed hook, show the exact source hook beside the new hook. Keep the sentence structure close, but change the concept completely.
13. The new idea must show the promised result in the first frame.
14. The viewer must leave with a useful explanation, prompt, skill, template, mini app or resource.
15. Make every tutorial usable by a non-coder unless I explicitly ask for code.
16. Reject creator-centred complaints, vague personal stories, generic listicles and hooks that require the viewer to know my private context.
FINAL OUTPUT
Start with a short audience section containing:
- who the audience is trying to become
- their five strongest observed needs
- exact example comments and the post links they came from
Then give me a creator section containing:
- each creator’s strongest posts
- the exact opening words
- the first-frame visual
- the format and editing pattern worth borrowing
- what appeared in winners versus ordinary posts
Then give me ten viable content ideas.
Finish with the best three. For each one include:
- exact source creator and post link
- exact source hook
- my translated hook
- first-frame visual
- bullet-point script with approximate timestamps
- what the viewer learns, does or keeps
- the reusable resource I should provide
- why it fits my audience and my skills
ACCEPTANCE CHECKS
Do not finish unless:
- the creator analysis comes from watched videos, not captions alone
- every audience claim includes exact comment evidence
- every recommended hook is traced to an exact source hook
- every recommendation gives the viewer a concrete outcome
- none of the final ideas simply repeats something I supplied
- all calculations were performed in Python
- missing evidence is labelled missing rather than invented
If any check fails, continue the research or return no final recommendations.That is the main prompt. You should not need to explain the entire project again after this.
What the agent should actually study
A useful creator breakdown is not a list of topics.
It should show:
- The exact words that opened the video
- The thing visible before the creator starts explaining
- How quickly the result appears
- The order in which information is revealed
- Where proof or credibility appears
- What the viewer is asked to do at the end
- Whether the same pattern appears in several strong posts
- What changes in ordinary or weaker posts
The agent is not copying a creator’s idea. It is learning the shape that made the idea easy to understand.
What the audience research should look like
Do not accept a made-up persona such as “Sarah, 27, loves matcha and wants to learn AI.”
Ask for exact questions from real comment sections.
Useful evidence sounds like:
- “Can I still try this if idk how to code?”
- “How does one get into coding? It sounds so fun but have no idea where to start.”
- “I still don’t understand what a cyberdeck is.”
Those comments immediately suggest useful content: beginner entry points, plain-English explanations and projects that feel exciting without requiring technical confidence first.
A comment proves that somebody asked the question. It does not prove that every woman in your audience feels the same way. Your agent should keep that distinction clear.
What a good hook translation looks like
A source video from @shebuildsrobots opened with:
“This is the stupidest thing I’ve ever made with my engineering degree.”
The useful pattern is:
- Show the absurd finished object immediately
- Contrast expertise with something silly
- Make the viewer wonder how and why it was made
A weak AI-generated version would be:
“Five fun ways to use AI.”
It loses the visual, the personality and the curiosity.
A close translation for my work is:
“This is the stupidest thing I’ve ever made with AI: a button that makes tech jargon speak English.”
Example script
First frame, 0–2 seconds
Show the working button turning an intimidating technical sentence into a normal definition, a concrete example and what the term does not mean.
Hook, 2–5 seconds
“This is the stupidest thing I’ve ever made with AI: a button that makes tech jargon speak English.”
Why it exists, 5–10 seconds
“I got tired of opening an AI tutorial and needing three more tabs just to understand the explanation.”
Quick demonstration, 10–22 seconds
Put three phrases through it: API, vector database and context window. Show the plain-English result after each one.
How I made it, 22–35 seconds
Explain the simple parts without teaching code: the instruction, the examples and the button that sends the text through it.
Payoff, 35–42 seconds
Show somebody using the same tool on a real work document or tutorial.
CTA, 42–46 seconds
“I put the exact skill and the browser version in the link so you can use it without building anything.”
What the viewer gets
A faster way to understand unfamiliar technical language, plus a reusable no-code tool.
That is a content idea. “Explain AI jargon” is only a topic.
If the results are still generic, send this
Reject the current shortlist.
You have summarised topics instead of transferring proven formats. For every idea, show me the exact source hook first. Keep its sentence structure and reveal pattern, then apply it to a completely different concept that fits my skills.
Remove any idea that:
- could be written without watching the source video
- has no visible result in the first frame
- gives the viewer no practical takeaway
- requires coding knowledge
- repeats an idea from my brief
- centres my feelings instead of the viewer’s problem
Return the revised ideas as:
SOURCE HOOK → NEW HOOK → FIRST FRAME → SCRIPT → VIEWER VALUE → RESOURCE.If the tutorials become too technical, send this
Rewrite these for a smart beginner who uses AI at work but does not code.
Do not teach syntax, terminals or software architecture. Let me handle the technical build behind the scenes.
The viewer should only need to:
- understand the problem
- see the result
- follow a few plain-language steps
- use the finished prompt, skill, template or appThe final idea format
Every recommendation should fit this simple block:
AUDIENCE QUESTION
The exact question or desire this answers.
SOURCE
Creator, post link and exact opening words.
MY VERSION
A close translation of the source hook using a new concept.
FIRST FRAME
The finished result the viewer sees immediately.
SCRIPT
Short bullet points in reveal order.
VIEWER VALUE
What she understands, does or keeps by the end.
RESOURCE
The prompt, skill, template, mini app or article she can save.If the agent cannot fill every line with real evidence, the idea is not ready.
The point is not to make content research sound technical. The point is to make it difficult for the agent to bluff.
Apify supplies the real posts and comments. The prompt forces the agent to watch, compare and cite them. /goal keeps it working until the output is genuinely useful.
Then the AI is no longer “coming up with ideas.” It is doing the research a good content strategist would do before recommending anything.