Mining Comments and DMs for Content Gold

Use only material you are entitled to review: comments on accounts you manage and messages sent to those accounts, subject to your organisation's policy. From another creator's public comments, record only a generalized topic and the post URL—never the commenter's identity or story. Do not automate collection or bypass platform controls.

Visibility is not consent. Ask before showing a username, quote, screenshot, voice note, order story, or identifying detail. DMs stay private; public commenters have not agreed to be featured. Without recorded permission, use only a generalized question.

Choose one handling path:

1. Minutes 0–2 — scope: choose one date window and one or two channels.
2. Minutes 2–10 — collect: scan your own surfaces manually for questions, objections, and repeated wording. On YouTube, use YouTube Studio > Community > Published; the official help says you can filter for questions. Confirm current menus on the official page.
3. Minutes 10–15 — redact locally: remove handles, profile links, order numbers, locations, faces, and contact details. Do not upload screenshots.
4. Minutes 15–18 — route: mark each row reply, content, ask_permission, or exclude.
5. Minutes 18–20 — cluster: group redacted rows; keep one-off questions as singletons.

Use this schema. Keep the source map separately; never put it in AI input.

Redaction is removal, not blurring. A screenshot can still expose a photo, phone, shop, timestamp, location, or distinctive wording. Transcribe the question, then check whether combined details identify the sender.

Before ChatGPT, review its official Data Controls. Turning off model improvement does not create consent. Manual spreadsheet clustering is a free alternative.

Paste only redacted records into this prompt:

Role: You are a qualitative research assistant.
Input: JSON lines containing signal_id, channel, redacted_question,
safe_phrase, permission, sensitivity, and path.

Tasks:
1. Flag any row that may still contain a person, phone, CNIC, address,
   order number, account handle, workplace, or uniquely identifying story.
2. Cluster the remaining rows by the need expressed, not by keywords alone.
3. Create a theme only when at least two distinct signal_ids support it.
4. Keep unsupported rows under "Singletons"; do not force a cluster.
5. For each theme, return: theme, supporting IDs, safe audience phrases,
   neutral question, suggested reply, and one content angle.

Rules:
- Do not infer age, gender, city, income, intent, or sentiment.
- Do not claim that "the audience" wants something.
- Never reconstruct removed details or produce a quote from a DM.
- Describe frequency only as "N of N collected signals in this date window."

<signals>
[PASTE REDACTED JSON LINES]
</signals>

A cluster is not a market conclusion. Write “delivery cost appeared in 3 of 20 collected signals from 12–18 July,” not “Pakistanis care most about delivery.” One post or group may dominate.

Preserve non-identifying phrases such as “delivery charges” for hooks, but paraphrase distinctive DM wording. Produce both a direct reply and a generalized content brief. If the value depends on a personal story, choose ask_permission.