HUVA academy
Free guides
Claude · Guide

Claude + Instagram: what actually works

AM Ani Mkrtchyan 18 min read Updated October 2026

You have seen the promise: connect Claude to Instagram, sit back, the page runs itself. I run 31+ agents, several for social media, and here is what is true: the connection is real, the sitting back is not. With a correct setup Claude reads your numbers, drafts your content and sorts your inbox. Your audience, your message and your taste still come from you, or you just produce the same slop faster.

In this guide
  1. The myths
  2. What an agent actually is
  3. How to connect: three tiers
  4. What Claude really does
  5. Control
  6. Where the hype starts
Part 1

The myths

MythIt grows the page by itself.
RealityClaude can tell you which of your last 30 posts beat your median and why. It cannot decide what you stand for. A page run on a generic brief posts generic content, engagement goes flat in two weeks, and the owner concludes "AI does not work for my niche". The brief was empty.
MythConnect it once and it posts daily.
RealityPosting is the one step most setups cannot do. The most popular free Instagram skill for Claude ends every workflow in a block you copy and post yourself. Real publishing needs Meta's official API (professional account, developer app, linked Page, long lived token, app review). Anything posting through a browser bot instead violates Instagram's Terms and gets the account action blocked.
MythAI finds viral content for you.
RealityIt finds what already went viral for others in your niche and ranks it against each account's own median. That is research, not prediction. The author of that skill tested its hook scorer on 74 real hooks: it reliably catches bad hooks, and between two decent ones it is a coin flip. The tool kills weak ideas. Your retention graph picks winners.
MythIt can reply to everyone like you.
RealityIt can draft a reply to everyone in a voice it learned from a file you wrote. Without the file you get polite, interchangeable answers that followers spot in one scroll. And drafting is not sending: auto DMs and mass replies through unofficial tools are exactly what Instagram restricts accounts for.
MythMore automation means more reach.
RealityMore automation means more output. Reach comes from how many people keep watching and how many save. Ten automated posts a day from one weak brief lower both, and the algorithm shows you less. In every account I have audited the fix was fewer posts with a sharper point of view.
Part 2

What an agent actually is

An agent is a new employee. It has instructions, keeps notes, can use certain tools, and has a boss who approves certain things. Give a new employee no brief and say "do our Instagram" and you get generic work, fast and polite. Same with Claude. Every part of the agent is a text file or a permission:

  • Voice file (taste). Who you talk to, what you believe, words you use and never use, why your best posts worked. The brief. Everything else reads it first.
  • Skills (procedures). Step by step instructions for one job: a Reel script, a caption check, comment sorting. A markdown file Claude follows.
  • Tools (what it can touch). A browser tab, an API connector, a local script. No tool, no action.
  • Memory (what it remembers). Your swipe file, last week's outliers, what you told it to stop doing.
  • The rule (what it cannot do alone). Publish, send, follow, delete. A line in the instructions plus a tool it does not have.

The clearest public example is the free, MIT licensed instagram-agent-skill on GitHub (3.4K stars): 13 skills, 5 local Python tools, one voice template. The whole "agent" is this:

Folder structure
instagram-agent-skill/
  skills/
    ig-reel/SKILL.md       one idea into a Reel: 3 hooks, script, beat sheet
    ig-viral/SKILL.md      what works in your niche, ranked by outlier multiple
    ig-caption/SKILL.md    caption linter, 125 character window, 5 hashtag cap
    ig-carousel/           ig-story/   ig-profile/   ig-plan/   ig-human/
    ig-comment/            ig-reply/   ig-dm/        ig-repurpose/  ig-audit/
  templates/
    voice.md               the taste file, empty until you fill it
  hookscore.py  beats.py  caption.py  humanize.py  detect.py  swipe.py
                           5 local tools, no network, nothing uploaded

installed on your machine:
~/.claude/skills/ig-*/        the procedures
~/.claude/instagram/voice.md  the brief every skill reads first
No part of this posts to Instagram. Every skill ends in a block you copy. The agent is the files; the connection is a separate decision.

Install: paste the repo URL into Claude and say "Install this skill, then confirm /ig-reel works". In Claude Code, git clone and copy skills/ig-* into ~/.claude/skills/. Without Claude Code, paste one SKILL.md at the top of a chat and it runs as a mode (you lose the Python tools). Then the first job, before anything else: fill voice.md, or send Claude three of your Reels and say "write my voice.md from these".

Part 3

How to connect: three tiers

Safest to most technical. Most readers should stop at tier 1 and stay there: you get analytics, drafts and inbox help today with zero account risk, because the only thing acting on Instagram is your own browser with you in front of it.

01Tier 1: Claude in Chrome. Start here.

Claude works inside your own logged in Instagram tab. No API, no token. You watch and approve every click, and you press publish.

Start here
Install Claude in Chrome, keep Instagram open in a logged in tab

That tab is the whole connection. Paste your voice.md at the top of the chat before any job.

Step 2
Read Insights

Open Professional dashboard and say: "Read reach, saves and shares for my last 30 posts from this tab and list the outliers above my median". Claude scrolls, reads, reports. Nothing changes on the account.

Step 3
Draft in the inbox and the composer

Open a DM or comment thread; Claude sorts and types a draft reply, then stops. Open the composer; Claude drops in the approved caption and hashtags, then stops. You press send and publish, every time. Scheduling needs tier 2.

02Tier 2: the official Meta API through a connector

Where scheduled publishing and analytics automation actually live. Claude gets real tools: pull insights, read comments, publish and schedule. The price is a real setup.

Step 1
Professional account linked to a Facebook Page

Business or Creator account connected to a Page you admin. Without this the API will not talk to you.

Step 2
A connector between Claude and Meta

I use Windsor.ai: it sits on the official Meta API, pulls Instagram organic insights per post, and has write actions for Instagram (image posts, comments) you switch on separately. Composio, Zapier, Make, Buffer or Metricool with MCP do the same job. You authorise Instagram inside the connector once (it handles the developer app, permissions and token), then add the connector to Claude.

Step 3
Read only for a week, then publish behind your "approved"

Run analytics and comment reading first, check the numbers against the Insights screen, then turn on scheduling. The rule stays: never call the publish tool without an explicit "approved" from you in the same conversation.

03Tier 3: approved DM partners

ChatPlace or ManyChat run keyword DM flows ("comment this word and I send you the guide") inside Instagram's rules. Claude sends nothing here. It writes the brain.

Step 1
Connect your account to the partner

The connection is theirs and sanctioned; that is the whole reason this tier exists.

Step 2
Claude writes the knowledge base and the reply scripts

Offers, prices, dates, FAQ, the keyword message, the follow up, the "a human will answer this" handoff. All drafted by Claude in your voice, all fired by the partner, only after someone writes first. Weekly, paste the unanswered threads back to Claude and it writes the additions.

Part 4

What Claude really does on your Instagram

Four jobs that work today, each with a prompt you can reuse. Every prompt opens the same way: role, voice file, data source, output format. Variables are in {{curly braces}}. The output comes back in whatever language your voice file says.

The voice file, filled in

The file all four jobs read, filled in for a Yerevan pastry shop I made up. Copy it, replace every line. The Best posts section with the "why" is the part most people skip and the part that teaches Claude your taste.

voice.md · worked example
# voice.md · Nar Pastry, Yerevan

## Who we are
Small pastry shop in Kentron, 2 locations, 6 people. Everything baked the same morning.

## Who we talk to
Women 25 to 45 in Yerevan who order for the office, a birthday or a Sunday table.
They care about freshness and whether it looks good on the table. They do not care about "artisanal".

## What we believe
* A cake that arrives at 9:00 was baked at 5:00, or it is not ours.
* Showing the kitchen beats describing it.
* One flavour done well beats a menu of twelve.

## Content pillars
1. Morning kitchen (process, hands, ovens, steam)
2. Orders and tables (what people ordered, how it looked)
3. Honest answers (prices, delivery, allergens, why we do not do fondant)

## How we sound
Short sentences. Armenian first, one English line only if the product name needs it.
Warm, a bit funny, never cute. We say "we", never "our team".
Words we use: fresh, this morning, Kentron, table, order by 18:00.
Words we never use: artisanal, premium, indulge, elevate, treat yourself, foodie.
Emoji: at most one per caption, usually none.

## Best posts (and why)
* Reel: 5:40 am, first tray out of the oven, no music, oven sound only. Saves 4x our median. Why: real, nobody talks, you can almost smell it.
* Carousel: "What 20,000 AMD buys for an office of 8". Shares 3x median. Why: a concrete number and a concrete table.

## Never
* Never invent a price, a date or an ingredient. Use {{price}} and flag it.
* Never promise delivery times we have not confirmed.
* Never reply to a complaint publicly before a human has read it.

## Calls to action we actually use
* "Order by 18:00 for tomorrow morning" (link in bio)
* "Which flavour should we bake on Saturday?" (comments)

01Analytics: find the outliers and explain them

Tier 1 (Claude reads Insights in your tab) or tier 2 (Windsor.ai pulls the numbers, no pasting). Skill version: /ig-audit.

Prompt · 30 post audit
Role: You are my Instagram analyst. You do not write content in this task, you explain numbers.
Voice file: use the attached voice.md as the definition of what "on brand" means for this account.
Data source: {{paste the last 30 posts as rows: date, format, hook or first line, reach, saves, shares, comments}} OR read them from the Insights screen open in my browser tab.
Output format:
1. My median reach, saves and shares across the 30 posts.
2. Every post above 1.5x median on any of the three, as a table: post, metric, multiple over median, format, pillar.
3. For each outlier, one paragraph on why, using only what is visible in the hook, format and pillar. Mark guesses as guesses.
4. Every post below 0.5x median reach, same table, same explanation.
5. Three patterns across the outliers that I can repeat next month, each tied to a pillar from voice.md.
Rules: never invent a number. If a value is missing, write {{missing}} and continue.
Ranking by multiple over your own median is what makes this useful. 8,000 reach is an outlier on a 2,000 median account and a flop on a 20,000 one.

02Content ideas: 20 ideas from your pillars and best 10

Tier 1, no connection needed. Skill versions: /ig-viral for your niche, /ig-reel to turn one idea into hooks, script and beat sheet.

Prompt · 20 ideas
Role: You are my content strategist. Your job is ideas with a point of view, not a content calendar.
Voice file: voice.md attached. The pillars, the beliefs and the "never" list are hard constraints.
Data source: my 10 best posts from the last 90 days: {{paste hook, format and why it worked for each}}.
Output format: a table with 20 rows: number, pillar, hook (under 12 words, from my voice), format (Reel / carousel / story / single image), the belief this idea challenges, the concrete scene or example it uses.
Rules:
* At least 12 of the 20 must challenge a belief my audience holds, not confirm it. Name the belief in plain words.
* No idea may reuse a hook structure from my 10 best more than twice.
* Nothing generic: every row must name a real scene from my business or a real question from my comments. If you do not have one, write {{need a real example}} instead of inventing.
* Flag any idea that would break a line in the "never" section.
The "belief it challenges" column is the filter. Ideas that only confirm what your audience thinks get polite likes and zero saves.

03Publishing: caption, first comment, 5 hashtags, linted

Tier 1 (Claude types it into the composer, you publish) or tier 2 (the connector schedules after your approval). Skill version: /ig-caption, which shows the 125 characters the feed displays before "... more" and enforces the 5 hashtag cap.

Prompt · caption package
Role: You are my caption editor. You write in my voice and you lint before you deliver.
Voice file: voice.md attached. Use its words list and its "never use" list literally.
Data source: the post: {{what the video or image shows, in 3 lines}}; the idea: {{the belief it challenges}}; the call to action: {{one CTA from voice.md}}.
Output format, in this order:
1. The first 125 characters, shown alone, exactly as the feed will cut them. They must contain the point, not the setup.
2. The full caption, under 900 characters, line breaks included, in {{language}}.
3. The first comment: context, the link mention or the question, under 300 characters.
4. Exactly 5 hashtags: 2 broad for my niche, 2 for my city or region, 1 for my own brand. No banned or generic tags.
5. A lint report: which "never use" words you removed, whether the 125 character window carries the point, character counts for each block.
Rules: no invented numbers or dates; use {{placeholder}} and flag. No emoji unless voice.md allows it.
Put the point in the first 125 characters, always.

04Inbox: sort, draft in your voice, flag what needs a human

Tier 1 (Claude reads the thread in your tab) or tier 2 (the connector reads comments). Skill versions: /ig-reply and /ig-dm.

Prompt · inbox triage
Role: You are my inbox assistant. You sort and you draft. You never send.
Voice file: voice.md attached. Replies must sound like the owner, not like support.
Data source: {{paste the comments and DMs from the last 24 hours, one per line, with the sender's first name}} OR read the open thread in my browser tab.
Output format: four lists, in this order.
1. LEAD: anyone asking about price, availability, booking or "how do I order". For each: one draft reply in my voice, under 60 words, ending with the next step from voice.md.
2. QUESTION: anyone asking about the product, process or an opinion. One draft reply each, answering directly, under 80 words.
3. SUPPORT: complaints, delays, refunds, anything emotional. Do not draft a public reply. Write a one line summary and mark NEEDS HUMAN.
4. NOISE: spam, bots, emoji only, "follow me". No reply. Just the count.
Rules: never invent a price, a date or a promise; write {{confirm}} where I must fill it in. Never draft a first message to anyone who did not write to me. If a message mentions health, legal issues or a public figure, move it to SUPPORT regardless of tone.
The SUPPORT bucket is the point. An agent answering an angry customer in a cheerful voice costs more than a day of unanswered DMs.
Part 5

Control

One rule: nothing goes out without your "approved"

Put it in the instructions: "Every output ends as a draft. Nothing is published, sent or scheduled without the word approved from me in this conversation". On tier 1 there is no publish tool anyway. On tier 2 keep write tools off until the read only week is done, and keep the rule after. Then the never list, at the bottom of voice.md:

voice.md · never section
## Never
* Never post, publish or schedule. Drafts only.
* Never send a first message to anyone. Reply drafts only, and only to people who wrote first.
* Never invent a number, price, date, ingredient or result. Write {{confirm}} and flag it.
* Never follow, unfollow, like in bulk or comment on other accounts without a named post and my approval.
* Never answer a complaint, a health question or a legal question. Mark NEEDS HUMAN.
* Never delete anything.
Every account restriction I have seen a client get came from breaking one of these with a tool, not with Claude.

Weekly check, 20 minutes

  1. Run the 30 post audit. Read the outliers and the flops.
  2. Update voice.md: a new "why" goes into Best posts; a line that produced a flop twice gets changed. The file changes when results change, not when you feel like it.
  3. Answer every SUPPORT item yourself.
  4. Pick three ideas for next week. Three, not ten.
Part 6

Where the hype starts

Inside Instagram's rules: your own browser with you in front of it, the official Meta API through an authorised connector, and approved DM partners that fire after someone writes first. Everything else is outside, and outside is where accounts get restricted.

⚠️

Outside the line: browser bots that log in as you, mass follow and unfollow tools, auto comment engines, auto DM tools that message people who never wrote to you, and anything sold as "AI that posts 10 times a day for you". Result: action block, then reduced reach, then a disabled account. Claude is not the risk. The tool holding your password is.

The test: does it need your Instagram password or a browser session it keeps without you? Wrong side of the line. Official API with a token you can revoke, or a tab you are looking at? Fine.

The videos promising "I connected Claude to Instagram and it 10x'd my page" show the connection, never the voice file. The connection took me an afternoon. The file took months of posting, reading the numbers and writing down why.

Summary

  • Connecting Claude to Instagram is real. Sitting back is not. The agent is only as good as the voice file you give it.
  • An agent is text files plus permissions: voice, skills, tools, memory, one rule. Not magic.
  • Tier 1, Claude in Chrome in your own tab, is right for 90% of readers. Tier 2, the official API through a connector, is the only place scheduling lives. Tier 3, approved DM partners, is the only sanctioned message automation.
  • Four jobs work today: audit, ideas, caption package, inbox triage. Four prompts above.
  • Browser bots, mass follow tools and auto DM engines are the hype, and the ban.
→

If you want to build the agent itself, with the voice file, the skills and the rule, we do it step by step in the HUVA Academy agents course. The rest of my setup, including what breaks, is on @aiwithani.

Related guides