How to get Claude or ChatGPT to analyse your Instagram reels and tell you what to post next
A general assistant knows a lot about content and nothing about your account. Ask what to post next and it will guess. The fix is to give it your real numbers, and there are three ways to do that.
Short answer. Claude and ChatGPT cannot read your Instagram account by themselves, so give them data in one of three ways. You can paste a table of your posts and insights, upload an export, or connect a tool through an MCP connector, which lets the assistant fetch social data while you chat. For XCut AI, add https://mcp.xcut.ai/claude/mcp as a custom connector in Claude, or add the server in ChatGPT. Then ask it which of your reels beat your average.
Why a plain chat cannot do this
Ask an assistant to review your reels with no data and one of two things happens. It tells you it cannot access your account, or it gives general advice dressed up as analysis. Even with web browsing, it often cannot read Instagram, since much of the platform sits behind a login and the app does not offer open access to other people's post data.
The assistant is good at the thinking part: spotting patterns, comparing posts, drafting. It only lacks the input. Each method below adds input at a different level of effort.
Method 1: Paste your numbers (free, 20 minutes)
This works today with any assistant and no extra tools.
- Open your professional account insights and list your last 20 to 30 reels.
- Note the date, topic, opening line, length, views, likes, saves, sends and follows. Use what your insights show.
- Put it in a table and paste it into the chat, or upload a spreadsheet.
- Ask a specific question, using the prompt below.
A prompt that works: "Here is a table of my last 25 reels. Work out my median views. List the posts at twice the median or more. Compare those with my bottom five on topic, opening line and length. Give me three hypotheses about what separates them, and quote the rows that support each one. Do not give general advice, and say plainly if there is too little data."
That last sentence matters. It tells the model it is allowed to say the data is thin, which cuts down on confident invention. Check its median yourself. Language models can slip on arithmetic unless they run code, so it is worth asking for the calculation to be shown.
The limit: you have only your own account, the data goes stale, and you did the typing.
Method 2: Add transcripts (free, a bit more work)
Numbers tell the assistant what happened. Transcripts help it see why. Add the spoken words of your top and bottom five reels. If you added captions, you already have them. Otherwise, transcribe them, or read them out and dictate. Then ask it to compare the openings and structures. You can do the same with a competitor's reels, as covered in how to compare competitor hooks.
Method 3: Connect it with MCP
MCP stands for Model Context Protocol. It is an open standard that lets an assistant call outside tools during a chat. In practice, you add a connector once, and the assistant can fetch data or take approved actions when your question needs it. You no longer paste tables.
XCut AI's connector gives an assistant access to social research. It can find outlier posts on a creator's or competitor's account, break down a pasted link into hook, structure and pacing, and write scripts and captions from that. It works with Instagram, TikTok, YouTube, Facebook and the Meta Ad Library.
Connect it to Claude
- In Claude, open Customize, then Connectors, then Add custom connector.
- Paste https://mcp.xcut.ai/claude/mcp.
- Hooks, playbooks and the quick tools work straight away. Tools that use your canvases, brand memory, imports or credits ask you to sign in to XCut AI the first time.
Connect it to ChatGPT
- Go to chatgpt.com/plugins, press +, and choose Add custom MCP server.
- Use the server address at https://mcp.xcut.ai/mcp.
- Approve the sign-in when asked.
App menus change often, so check the connector page for current steps. It also covers Gemini, Grok, Cursor, VS Code and others.
What to ask once connected
- "Which of my reels from the last few months beat my average, and what do they share?"
- "Find the outliers on these three competitor accounts."
- "Break down this reel: [link]. What is the hook and why does it work?"
- "Compare my best reel with this competitor's best one."
- "Write a script for my next reel using the strongest pattern, in my voice."
- "Turn those findings into a posting plan for the next two weeks."
Permissions and safety
Be careful about what any connector can touch. Read what it asks for when you approve it. XCut AI uses OAuth with two scopes, one to read and one to write inside your XCut workspace. It only asks for what the tools you use need, and it does not read your ChatGPT or Claude chat history. It does not post to social media or spend money. It researches, writes and plans. Your Instagram password is never involved, because it works from public posts, and it does not connect to your private insights.
That last point is a real limit. If you want saves, sends and follows in the analysis, you still need Method 1 for those numbers. A good workflow is to combine them: paste your private insights and let the connector fetch public competitor data.
Copy-and-paste prompts for Method 1
These work with a table of your own numbers pasted into any assistant. Run them one at a time, because a single large request tends to produce a vague answer.
Find what stands out
"Here are my last 30 reels with date, topic, length, views, saves and shares. Work out the median views. List every reel at twice the median or more, and every reel at half or less. Do not explain why yet."
Compare the top and the bottom
"Compare the top five with the bottom five. List only differences you can see in this data: topic, length, day of the week. For each one, say how many of the five show it. If the sample is too small to tell, say so."
Turn it into a test
"Based on that, write one sentence saying what you think helped, and how my next three posts could test it. Tell me what result would prove you wrong."
Asking it to name what would prove it wrong is the most useful line here. Assistants tend to agree with whatever you hand them, and this line forces a claim you can check.
How to build the table in about 15 minutes
- Open your professional dashboard and go to your content insights. Exact menus change, so look for the list of recent reels.
- Create a sheet with these columns: date, link, topic in three words, length, views, likes, comments, saves, shares, follows.
- Fill in your last 20 to 30 reels. Leave out anything from the last few days, since those counts are still moving.
- Copy the rows and paste them in as plain text.
- Save the sheet. Next month you only add new rows.
Writing the topic yourself takes the longest, and it is worth doing. The assistant cannot see your video from a table, so without that column it has only the numbers to go on.
Mistakes that give you confident nonsense
A general assistant will always give you an answer, including when your data cannot support one. One creator gave this advice about YouTube: "Run your thumbnail, title and description through ChatGPT and do what it tells you". It can be a helpful second opinion. Doing what it tells you, without checking, is where the trouble starts. Watch for these.
- It claims to have looked at your account. A plain chat with no connection or browsing cannot see it. If it talks about your posts without being given them, it is guessing.
- It quotes algorithm rules with total confidence. Ask for a source. If it cannot give one you can open, treat it as an opinion.
- It finds a pattern in three posts. Nothing can be concluded from that. Ask how many posts support each claim.
- It says Tuesdays are your best day. Day-of-week results from 30 posts are mostly noise. Do not rebuild your schedule on them.
- It gets the maths wrong. Check the median yourself, or ask it to show the sorted list. A sheet formula is safer.
- It flatters you. Ask it to criticise your weakest reel as a strict editor would.
Claude or ChatGPT: does it matter?
For this job, the method matters more than the model. Both can read a table, compare posts and draft a script. The differences that matter in practice are what your plan allows, such as file uploads, projects and connectors, and which one you already pay for. Features and limits change often, so check the current plans. Run the same prompt in both, using the same table, and see which answer is more useful for your account. People use these tools in many ways. One commenter wrote: "Chatgpt helps sooo much tbh, i always ask him for a lot of ideas and to help fix my scripts". Another said: "Claude is good for writing blogs". Your own test will tell you more than either opinion.
A monthly routine that takes 30 minutes
- Add the month's reels to your sheet.
- Run the three prompts above.
- Compare the answer with last month's, and note whether your test worked.
- Check two competitors for new outliers, by hand or with a connector.
- Pick one thing to test next month, and write it at the top of the sheet.
Keeping everything in one project or thread helps, since the assistant can refer to earlier months. Keep your sheet as the master copy, though. Chats get long and details get lost.
Which method should you use?
| Paste a table | Add transcripts | MCP connector | |
|---|---|---|---|
| Cost | Free | Free | Free tools, then credits on a plan |
| Setup | None | None | A few minutes, once |
| Your private insights | Yes | Yes | No |
| Competitor data | By hand | By hand | Yes |
| Stays current | No | No | Yes |
| Best for | A first look | Studying hooks | Regular research |
Getting better answers from any method
- Ask for evidence. "Quote the posts behind each claim." It makes vague answers easy to spot.
- Ask for a comparison. Top against bottom works better than "analyse this."
- Give your goal. Followers, leads and views lead to different advice.
- Limit it. Ask for three ideas you can test this week and not a full strategy.
- Save your context. A project with your voice file means you do not start over. See how to make AI scripts sound like you.
- Remember the sample. Five posts cannot prove anything. Treat answers as things to test.
If you only want a first look without setting up anything, the free scan shows which of your public posts beat your average. Then you can bring that list into the chat of your choice.
Frequently asked questions
Can ChatGPT analyze my Instagram account?
Not by itself. It can analyse data you paste or upload, or data fetched by a connector such as XCut AI. Without that it can only give general advice.
What is an MCP connector?
A way for an assistant to call outside tools during a chat, based on the open Model Context Protocol. You add it once, and the assistant can fetch data when needed.
Do I need to give my Instagram password?
No. XCut AI works from publicly available posts and never asks for your Instagram password. Be wary of any tool that does.
Can a connector see my private Instagram insights?
XCut AI cannot. It does not connect to your analytics, so saves, sends and follows have to be pasted from your insights if you want them in the analysis.
Is it free?
Hooks, playbooks and the quick tools work without an account. Tools that use imports or credits need an XCut AI plan, shown on the pricing page.
Can I upload a screenshot of my insights instead of typing the numbers?
Often, yes. Most assistants can read images on plans that allow uploads. Check the numbers it pulled out before you trust the analysis, because small misreads change the results.
Why does the assistant give different answers each time?
These models do not give identical output on every run. Use them to raise ideas to test, and rely on your sheet for the numbers.
References
- XCut AI inside ChatGPT, Claude and other apps.
- XCut AI developer documentation.
- Model Context Protocol.
- XCut AI pricing.
See which of your posts beat your average
Paste your handle at xcut.ai/instagram for a free scan with no account, or start a trial to research any niche.
Start your 7-day trial$5 for 7 days · 1,000 credits · cancel anytime