reels-analyzer

Scrape competitors' Instagram reels, transcribe the audio verbatim, and use Claude to return concrete reel scripts you should shoot next week.

Open source under AGPL-3.0. The managed version lives inside The Automation Founders community at buildwithsumit.com.

What it does

You give it:

• your Instagram handle,
• your website URL (we read it for you — no need to write a business pitch),
• 1-5 competitor handles you want to study.

It runs a five-stage pipeline:

1. Read your site — fetches your homepage, strips the HTML, and asks Claude to write a one-paragraph business-context (what you sell, who the buyer is, your angle). Stored once per project.
2. Scrape — pulls the latest reels from each handle via Apify's Instagram Reel Scraper actor (caption, view/like/comment counts, video & audio URLs).
3. Transcribe — downloads the audio of each top reel and runs faster-whisper locally to get verbatim spoken text. Captions only tell you what was written; transcripts give you the actual hook the creator used in the first 3 seconds.
4. Analyze — feeds everything (your business context + your reels + each competitor's top reels with transcripts and view counts) to Claude Sonnet with a system prompt designed to be brutally specific.
5. Output — a one-page markdown report ending with exactly 5 numbered reel scripts you can shoot this week, each with a verbatim hook, 3-5 sentence body, and a CTA in your voice.

Why this exists

Most "AI content tools" stop at "write me a reel script about X." That produces generic creator slop. This tool grounds Claude's output in real high-performing reels from people you already think of as competitors — with the actual spoken words, not just the captions — and forces Claude to quote the evidence. The result reads like a research note from an analyst who watched 30 reels and tells you exactly what the top performers do differently.

What it's not

• Not a generic IG analytics tool. It doesn't track follower growth or run daily reports — that's what Apify or HypeAuditor are for.
• Not an autoposting tool. It returns scripts; you shoot them.
• Not a competitor to ChatGPT/Claude-as-a-coach. It's a narrow, opinionated pipeline that gives one specific kind of output well.

Architecture

The pipeline, end-to-end:

Step	Tool	What it does	Where it lands	Cache
Setup	Claude Sonnet (via Anthropic API)	Reads your website (HTML-stripped) and writes a one-paragraph business context: what you sell, who the buyer is, your angle	reels_profiles.business_context	Per project
1	Apify (instagram-reel-scraper)	Pulls latest reels per handle: caption, view/like/comment counts, video & audio URLs	reels_cache	24h per handle
2	faster-whisper (base model, int8 CPU)	Downloads each top reel's audio, transcribes the spoken words verbatim	reels_transcripts	Permanent (keyed by shortcode)
3	Claude Sonnet (via Anthropic API)	Reads business context + member reels + competitor reels with transcripts, writes a one-page markdown report ending in 5 ready-to-shoot scripts	reels_reports	—
4	Claude-aesthetic HTML (built-in)	Optional: renders setup form, dashboard, and report viewer for host apps	—	—

The pipeline runs as a background thread spawned by start_report_async(), so the HTTP request that triggers it returns immediately. The report row flips through pending → running → done|failed, with status_detail updated at each phase ("scraping 2 handles via apify", "transcribing reels via whisper", "asking claude to synthesize"). The Claude-aesthetic report viewer auto-refreshes every 6 seconds while the run is in flight.

Why each tool

Why Apify (not yt-dlp / instaloader). Instagram aggressively blocks anonymous scraping in 2025+ — 403 Forbidden on the GraphQL endpoint, even for public profiles. Free tools require a logged-in session cookie (your real account gets flagged) or a throwaway account (bans within weeks). Apify maintains a hardened actor with rotating residential proxies and absorbs the ban risk on their side. Pricing is ~$0.01-0.03 per profile scrape — negligible at a handful of handles per member per week.

Why faster-whisper (not the OpenAI Whisper API). Free, local, no API key, no audio data leaving the box. The base model is ~140MB, runs at int8 quantization on CPU, and processes ~3-5× realtime on a 4-core machine — so a 60-second reel transcribes in ~15 seconds. Quality is plenty for hook extraction (we're not transcribing legal depositions). First call lazy-loads the model into memory (~1 GB resident); subsequent runs in the same process reuse it.

Why Claude Sonnet (not GPT-4 / smaller models). Quality matters more than cost on the analyze step — a mediocre report makes the whole pipeline pointless. Sonnet follows complex system prompts faithfully (the 5-script-with-HOOK/BODY/CTA structure), is honest about ambiguity ("the data here is too thin to tell"), and won't slop out generic creator advice. About $0.05-0.10 per report at typical sizes (~15k input + 2k output tokens).

Why MySQL (not SQLite / Postgres / a vector store). Boringly available everywhere; the host app probably already has one. No vector search needed because Claude reads the full context per run — we're not retrieving similar past reports, we're regenerating fresh each time. The schema is five flat tables, in migrations/001_init.sql.

Stack

• Python 3.10+
• MySQL 8+ (or compatible — MariaDB works) for the report + cache schema
• ffmpeg (required by faster-whisper for audio decode)
• Apify account + token (free tier covers ~$5/month of scraping — plenty for a handful of handles per week)
• Anthropic API key for Claude Sonnet (roughly $0.05-0.10 per report at the volumes this tool runs at)

Install

pip install git+https://github.com/Build-With-Sumit/reels-analyzer.git

Or for local development:

git clone https://github.com/Build-With-Sumit/reels-analyzer.git
cd reels-analyzer
pip install -e .

You also need ffmpeg system-wide:

# Ubuntu/Debian
sudo apt install ffmpeg

# macOS
brew install ffmpeg

# Windows (via winget)
winget install Gyan.FFmpeg

Setup

1. Create the schema in your MySQL database (apply each migration in order):

   mysql -u<user> -p<pw> <dbname> < migrations/001_init.sql
   mysql -u<user> -p<pw> <dbname> < migrations/002_website_url.sql

2. Set environment variables (or put them in a config MySQL table — same keys; DB values win over env):

   export DB_HOST=127.0.0.1
   export DB_PORT=3306
   export DB_USER=youruser
   export DB_PASSWORD=yourpassword
   export DB_NAME=yourdb
   export APIFY_API_TOKEN=apify_api_xxx       # from apify.com/account/integrations
   export ANTHROPIC_API_KEY=sk-ant-api03-xxx  # from console.anthropic.com

Use it as a library

import reels_analyzer as ra

email = "[email protected]"

# 1) Auto-derive business context from the member's website (~10-15s)
business_context = ra.summarize_website("https://acme.com")

# 2) Save the project (profile + tracked handles)
ra.upsert_profile(
    email,
    ig_handle="acmefounder",
    business_context=business_context,
    website_url="https://acme.com",
)
ra.set_handles(email,
    self_handles=["acmefounder"],
    competitor_handles=["lukebuildsai", "gregisenberg", "wowxmanish"])

# 3) Kick off a background report run
report_id = ra.start_report_async(email)

# 3) Poll until done
import time
while True:
    r = ra.get_report(report_id, email)
    print(r["status"], "—", r["status_detail"])
    if r["status"] in ("done", "failed"):
        break
    time.sleep(5)

# 4) Read the markdown report
print(r["body"])

First run takes 3-8 minutes (scrape + transcribe). Subsequent runs are faster because both the scrape and the transcripts are cached for 24h.

Embed the UI in your own server

The package ships Claude-aesthetic HTML pages you can drop straight into any Python HTTP server (http.server, FastAPI, Flask, Django — anything). They return full HTML documents, not fragments:

import reels_analyzer as ra

profile = ra.get_profile(email)
handles = ra.list_handles(email)
reports = ra.list_reports(email)

html = ra.dashboard_html(email, profile, handles, reports,
    run_url="/app/reels/run",
    setup_url="/app/reels/setup",
    report_url="/app/reels/report",
    back_url="/")
# self._send_html(200, html)

Three pages: setup_html(), dashboard_html(), report_html(). Each takes URL overrides so you can mount the feature anywhere.

Cost per report

At the default settings (15 reels per handle, 10 transcribed):

• Apify: ~$0.01-0.03 per profile scraped
• Whisper: $0 (runs locally on CPU; ~3-5 min for 50 reels combined)
• Claude Sonnet: ~$0.05-0.10 per report (~15k input + 2k output tokens)

So a report covering 5 competitors costs roughly $0.10 - $0.25 to run once and is cached for 24h.

Don't want to host it?

The managed version is inside The Automation Founders — $99/month, no infra, runs weekly automatically, plus other AI/automation skills Sumit ships every week.

License

AGPL-3.0-or-later. You can run it for yourself or your own business as much as you want, for free. If you run it as a network service for other people, you must publish your modifications under AGPL too — that's the part of the license that funds the open-source work.

Want to run a hosted version for your customers without that obligation? Contact [email protected] about a commercial license.

Contributing

Issues and PRs welcome at github.com/Build-With-Sumit/reels-analyzer.