Verified against Grok · 2026-07-22
Turn live X mentions into a brand sentiment scan Grok can actually see
Pulls a real-time read on how a brand, product, or campaign is landing on X right now, weighting recency and separating organic reaction from bot and engagement-farming noise, instead of a static analytics-dashboard sentiment score.
The prompt
Ready to copy — highlighted parts are example details you can swap.
You are running a real-time sentiment scan using Grok's live access to the X post index — not a generic social-listening report written from memory or a stale training snapshot. Every claim about "what people are saying" must trace back to posts from the stated time window, not general impressions of how a brand like this is usually discussed. SUBJECT The DTC skincare brand Lumenne, specifically its new retinol serum launch TIME WINDOW Past 6 hours, since the launch post went up at 9am ET WHY THIS SCAN MATTERS RIGHT NOW Just launched a paid influencer push this morning; marketing wants to know if it's landing before doubling down on spend this afternoon. KNOWN NOISE SOURCES The brand's own affiliate program pays a flat fee per post using the hashtag #LumenneGlow, so a wave of near-identical praise is expected and should not count as organic sentiment. SEGMENTS TO BREAK OUT Paid affiliate posts, unpaid customer reactions, and skincare-community accounts known for ingredient-level critique SCAN RULES Search the live X post index for The DTC skincare brand Lumenne, specifically its new retinol serum launch across Past 6 hours, since the launch post went up at 9am ET, not your training data's general sense of the brand's reputation — if a claim can't be tied to an actual post from this window, it does not belong in this scan. Separate genuine organic reaction from engagement-farming and bot amplification: a cluster of near-identical phrasing posted within minutes of each other, accounts with no history beyond this topic, or reply-guy quote-tweets clearly written to ride a trending tag are noise, not sentiment, and inflating a sentiment score with them misrepresents what real users think. Weight recency deliberately — a post from ten minutes ago tells you something a post from three days ago inside the same window does not, especially if sentiment is visibly shifting within the window itself; call out a shift in direction if one is happening, not just a net average that hides it. Break sentiment out by the segments named above rather than a single blended number, since a single average can hide an angry vocal minority sitting inside an otherwise neutral majority. Quote actual posts, paraphrased or lightly redacted for anonymity if the account is not a public figure or verified brand account, as evidence for every claim — a sentiment finding with no quoted post backing it is an assertion, not a scan result. Distinguish a genuine complaint about the product or brand from a complaint about something adjacent it is getting blamed for unfairly — a shipping delay caused by a carrier, a feature request being read as a bug — and flag the distinction rather than folding both into one negative bucket. If the volume of relevant posts in this window is too thin to support a real read — a few dozen posts, mostly from accounts that also posted about ten other brands today — say so plainly instead of manufacturing a confident sentiment score from noise-level volume. OUTPUT FORMAT 1. Headline read: one sentence, net direction and whether it's stable or shifting within the window. 2. A table: Segment | Sentiment (positive / mixed / negative) | Approximate volume | Representative quote | Noise filtered out. 3. Anything you excluded as bot or engagement-farming noise and why, specific enough that someone could spot-check your call. 4. A confidence note: how much real signal this window actually contained, not just the sentiment conclusion itself.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Grok's advantage here is native indexing of X's live firehose, so it can retrieve posts from minutes ago rather than being limited to a training-data cutoff or a periodic search-engine crawl; this is different from a generic LLM asked to guess sentiment from memory of how a brand is generally discussed, which produces a plausible-sounding but time-blind answer regardless of what's actually happening right now. The instruction to distinguish engagement-farming and bot clusters from organic reaction matters because live social platforms have real coordinated-posting patterns triggered by affiliate programs, brand-ambassador rewards, and clout-chasing quote-tweets, and a naive volume-based percent-positive score treats a paid coordinated wave identically to unprompted reaction, producing a number that's technically accurate about what got posted and completely wrong about what it means. Requiring quoted-post evidence for every claim converts sentiment analysis from an unfalsifiable vibe into something checkable — a stakeholder can pull the same window and verify the read, which is the only way a live-social sentiment claim earns any trust at all. Breaking sentiment out by segment rather than one blended number addresses a specific statistical trap: an angry, vocal minority of ingredient-critique accounts can sit inside an otherwise neutral-to-positive majority, and a single averaged score would either wash that minority out entirely or let it drag the whole number down in a way that misrepresents both groups. The explicit low-volume honesty rule matters because a six-hour window on a mid-size brand can easily produce sentiment-adjacent noise dressed up as signal, and a model under pressure to 'give a read' will smooth over thin data into false confidence unless directly told that naming the thinness is itself part of the job, not a failure to complete it. Weighting recency and explicitly calling out a shift in direction within the window closes a gap a single net score always creates: two scans that both land on 'mostly positive' can describe completely different situations if one is steady across six hours and the other flipped from strongly positive to increasingly negative in the last ninety minutes, and only the second version tells the marketing team the thing they actually need before deciding whether to keep spending into the same audience this afternoon.
What you get back
Headline read: net positive but softening — the first two hours skewed strongly positive on unboxing photos, but the last ninety minutes show a rising thread of "why does this smell like the old formula" replies. Segment | Sentiment | Volume | Quote | Noise filtered Paid affiliate | positive | ~140 posts | "obsessed with the new serum, #LumenneGlow" | Excluded from organic read — flat-fee affiliate copy Unpaid customers | mixed, trending negative | ~35 posts | "reformulated and nobody told us?" | none Ingredient-critique accounts | negative | ~9 posts | breakdown thread flagging a changed preservative | none, small but credible sample Filtered: 40+ near-identical "can't wait to try this" posts from accounts created this week, consistent with a giveaway-entry requirement, not genuine anticipation. Confidence: moderate on the positive affiliate read, low-but-real on the reformulation complaint — only 9 accounts, but all from established skincare-critique voices, worth escalating before the afternoon spend decision.
Verified against
Grok Grok 4.1 · 2026-07-22
Changelog
- 2026-07-22 — Initial publish, verified against Grok 4.1 with live X index access.
Need this built into your business?
If a prompt isn't enough — what Scult builds, built and maintained for you — that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
