Build your own Shorts Scraper

We are starting Gate 1 of my Shorts Scraper build.

Interview me before recommending creators or tools. Ask no more than five questions at a time and wait for each answer.

You need to learn:
- my niche and sub-niches
- the audience I want to attract
- the product, service, or next action my content supports
- the platforms and regions I care about
- the content formats I can realistically make
- the result I want this research library to help me improve
- the claims or topics I will not cover
- my available time, budget, and comfort with technical setup

Do not generate content ideas yet. Do not recommend creators yet. When the interview is complete, write a concise Research Brief with these headings: Audience, Business Goal, Content Scope, Formats, Platforms, Evidence Standard, Constraints, and Questions This Library Must Answer.

Then ask me to approve or correct the brief before moving to Gate 2.

We are starting Gate 2. Use my approved Research Brief.

Help me build a watchlist of 5 to 15 creators. First ask whether I already have saved posts, creator handles, or competitor research. If I do, inspect those before suggesting anyone new.

For every creator, record:
- exact display name and handle
- platform
- why this creator is relevant to my audience or format
- the specific trait I want to study
- whether the identity is verified
- whether they are a direct competitor, adjacent creator, or format reference

Reject accounts that are a poor audience match, inactive, misidentified, mostly sponsored, or impossible to verify.

Use this exact outlier protocol:
- Under 200,000 followers: a candidate qualifies when candidate views are at least 5 times the creator's follower count OR at least 5 times the median views of the creator's latest 10 videos.
- At 200,000 followers or more: ignore the follower ratio and require candidate views to be at least 2 times the median views of the creator's latest 10 videos.

For every creator, preserve the follower count and capture time. For every median test, preserve the 10 individual baseline view counts, sort them, and show the median calculation. Never include the candidate video in its own baseline. If the follower count, candidate views, or required 10-video baseline is unavailable, label the result insufficient evidence. The 5x and 2x thresholds are research filters, not virality guarantees.

Return a Creator Watchlist table and a Collection Plan. Stop for my approval before Gate 3.

We are starting Gate 3. Audit the tools and access available in this workspace before touching any social URL.

Check for:
- a file-capable AI agent such as Claude Code or Codex
- a folder where source videos and records can live
- FFmpeg and ffprobe for media inspection and frame extraction
- a transcription option such as local Whisper
- a browser or approved downloader for sources I am allowed to collect
- a spreadsheet, CSV, or JSON store
- a local web framework and a hosting option
- an optional collector such as the Scrape Creators API or another source that passes a live test

Use ChatGPT Sites when this task runs in ChatGPT with Sites access. Keep the Site owner-only, plus workspace administrators when applicable. If this task runs in Claude Code, state clearly that Claude Code cannot create a ChatGPT Site. Propose a separate host and authentication method, and wait for my approval before deploying it.

Create four lists: Available Now, Requires Installation, Requires Login or Payment, and Must Be Done Manually.

Rules:
- Do not ask me to paste passwords, cookies, or API keys into chat.
- Do not install software, spend money, log in, upload files, or deploy without asking at the moment that action is needed.
- A tool claiming support for Instagram or TikTok is not proof that it works today. Require a live test.
- A downloader working is not permission to republish someone else's media.
- If you cannot retrieve a video, ask me to supply the file. Do not fabricate a transcript or analysis.
- Do not add user accounts, a user database, or a multi-tenant backend. This version has one owner and one editable creator configuration.

Recommend the smallest manual stack that can complete one ten-video proof. Stop for approval before Gate 4.

We are starting Gate 4. Help me collect the requested daily batch of 1 to 10 videos from the approved creator set.

Create one source receipt per video with these fields:
- research ID
- creator and verified handle
- platform and source URL
- date posted, if visible
- date and time captured
- views, likes, comments, shares, and saves exactly as visible
- creator follower count and capture time
- the view counts for the creator's latest 10 baseline videos when the protocol requires them
- the sorted baseline values and calculated median
- candidate views divided by follower count
- candidate views divided by the latest-10 median
- the exact qualification rule that passed
- caption text
- local video filename and file size
- acquisition method
- missing fields and why they are missing
- rights or usage note

Preserve the original files. Create smaller playback copies only when the site needs them. Never overwrite the archive.

After each collection, verify that the URL, creator, file, duration, orientation, and visible metrics belong to the same post. Exclude the candidate from its own latest-10 baseline. Leave unavailable metrics blank. Do not convert blanks to zero.

Search the newest 30-day window first. When all unseen outlier options there are exhausted, continue through older non-overlapping 30-day windows until the requested batch is full, the source history ends, or the approved spending cap is reached. Deduplicate against every prior result before further processing. Fill the batch with verified outliers first. If fewer outliers exist than requested, use the strongest unseen remaining videos and label each one "best available, not a verified outlier." Reject any video that cannot be played inside the review page or transcribed. Return fewer only when no unseen playable, transcribable candidates remain. Do not analyze patterns yet. Stop before Gate 5.

We are starting Gate 5. Work through every selected source receipt without asking me which videos to process.

For each video:
1. Transcribe it with timestamps.
2. Correct obvious transcription errors only when the audio or on-screen text supports the correction.
3. Extract representative frames at the opening and every meaningful visual change.
4. Record the exact spoken hook, on-screen text hook, and visual hook.
5. Map the structure beat by beat.
6. Identify the topic, angle, promise, value, proof, CTA, editing format, editing style, pacing, pattern interrupts, and ending.
7. Separate observation from interpretation. Quote only short phrases needed to support the analysis.
8. Mark anything you could not inspect as unknown.

Do not copy a creator's whole script or identity. Focus on reusable decisions such as a hook shape, sequence, proof placement, or visual format.

Confirm that each video plays inside the review page through an approved embed or permitted local playback copy. Return one standardized analysis record per video. Do not reveal the batch until every selected video is playable, transcribed, analyzed, and tied to its source receipt. Replace any record that fails those checks with the next eligible candidate.

We are starting Gate 6. Compare only the ten verified analyses and the approved Research Brief.

Build a Pattern Report with:
- hook patterns that repeat
- common structures and their order
- recurring angles
- proof types and where proof appears
- CTA patterns
- editing formats and visual changes
- pacing ranges
- audience problems and desires that recur
- patterns that fit my realistic production constraints
- one-off choices that should not be treated as rules
- missing evidence that prevents a conclusion

For every conclusion, cite the research IDs that support it. Separate Frequency, Possible Function, and Confidence. Do not claim causation from views alone.

Then produce a Modeling Menu. For each promising pattern, state what can be borrowed ethically, what must come from my own experience, and what would make the result too close to the source.

Do not generate finished content ideas unless I ask after reviewing the report. Stop for my corrections before Gate 7.

We are starting Gate 7. Build the smallest private research site that makes the approved evidence easy to study.

Preferred route: create and deploy an owner-only ChatGPT Site when this task runs in ChatGPT with Sites access. Verify the access setting after deployment. If this task runs in Claude Code, do not describe the result as a ChatGPT Site. Show me the separate hosting and authentication plan and wait for approval before deployment.

Required experience:
- creator filter
- search across hooks, transcripts, captions, and structures
- embedded or permitted local video playback
- transcript, caption, and analysis tabs
- captured metrics with the capture date
- follower count, latest-10 median, calculated ratios, and qualification rule
- original source link
- previous and next navigation
- a visible research ID
- explicit unknown or insufficient baseline labels
- a saved integer control labeled "Videos per day" with a minimum of 1 and maximum of 10
- a button labeled "Get the next batch"
- a visible queued reveal date and time
- progress, last-run time, records checked, records accepted, records rejected, cost, and errors for every button run

Use my brand colors and fonts, but prioritize legibility. Do not copy another site's logo, product name, or exact visual identity. Do not expose private files, credentials, cookies, or local paths.

The button is the only collection trigger. When I click it, use the saved daily batch size, collect a fresh candidate pool from the approved creator list, and deduplicate by platform and post ID or canonical URL. Search the newest 30-day window first. When its unseen outliers are exhausted, move through older non-overlapping 30-day windows without recycling videos. Fill the batch with qualified outliers first. If fewer outliers exist than requested, use the strongest unseen remaining videos and label each one "best available, not a verified outlier."

Prepare every selected record before revealing it. Each record needs working playback inside the page, a timestamped transcript, completed analysis, and its source receipt. Replace any candidate that fails playback or transcription. Queue the complete batch for the next local calendar day at the saved review time. Use an earlier date or time only when I explicitly request it for that batch.

Do not add an automatic recurring daily run, morning notifications, email alerts, phone alerts, user accounts, or a database of users. Store one owner's editable creator list, daily batch size, review time, and settings in a simple configuration file. Keep API secrets in the hosting environment, never in browser code or the repository.

Build from the verified manifest instead of hand-entering records twice. Add checks that fail when a video, transcript, receipt, or required analysis field is missing. Verify every selected record, mobile layout, keyboard access, playback, copy buttons, source links, search, daily batch-size bounds, queued reveal time, and next-day release behavior.

Deploy through the approved path or keep it local. Report what is real, what is static, and what is still manual. Do not describe the button as working until one live click retrieves, filters, saves, prepares, queues, and reveals at least one record at its scheduled time.