crowd-positioning

name

crowd-positioning

description

Read how the crowd is POSITIONED around a coin during a social spike and whether that positioning is EXTREME vs the coin's own history — not a news summary. Use when a token is trending, its social volume / dominance is spiking (now or at some past date), or someone asks "what happened with X / why is it being talked about / is this hype real / is the crowd long or short here / who is driving it". Turns raw `social_messages` into four signals: (1) extremeness vs a trailing baseline (percentiles, not adjectives), (2) organic vs manufactured (copies, channel/author concentration, acceleration shape), (3) narrative-vs-chain divergence (the crowd's concrete claims tested against on-chain flows), (4) crowd price-level clusters (the support / targets people actually name). Surfaces the dominant trending topics driving the spike (the "what happened") and delivers a positioning VERDICT with denominators and source links, then answers follow-up questions on the same pulled data.

Crowd positioning

You are positioning the crowd, not reporting on it. The failure mode this skill exists to kill is journalism with the data — a readable recap of what people said. Nobody pays for the recap. They pay for the answer to one question:

How is the crowd positioned, and is that positioning extreme?

Every line you emit is a signal with a baseline, never a story. Not "lots of bullish ETF talk" — that's journalism. Instead: "social volume is 94th percentile of trailing 90d; 71% bullish vs a 53% trailing-30d median; 68% organic and still accelerating; chain confirms it (–$210M net off exchanges this window)." If a sentence has no number and no comparison, cut it.

When to use

A token is trending or its social dominance/volume spiked (now or a named past window) and the user wants to understand it — "what happened with $X around ", "why is everyone talking about Y", "is this real or bots", "is the crowd long here". You have a token and (ideally) a time window. If the window is missing, infer the spike window first (trending_stories / fetch_metric_data on social volume to find the peak) and state the window you chose.

The four signals (this is the whole product)

1. Extreme vs history? — A spike is meaningless without its own baseline. "12th percentile of trailing 90d" is a signal; "high social activity" is not. Rank this window's social volume, dominance, and sentiment skew against the coin's trailing distribution.
2. Organic or manufactured? — Same volume means opposite things if it's 5k people or one room pasting 200×. Use the dedup copies count, channel/author concentration, and the acceleration shape. Output a number: "70% organic, still accelerating" vs "bot campaign, 3 channels = 80%".
3. Does the narrative match the chain? — The crowd's claims are testable. "BlackRock accumulating" + chain shows $150M flowing to Coinbase = a divergence worth flagging. Pull each checkable claim's on-chain metric and report match or divergence, with the number.
4. Where does the crowd put price? — Price levels named in messages are positioning data. Extract them, cluster them, and show where the crowd places support / current / targets, with how many voices back each level. (E.g. "62772 was a trend word" → a level, not a topic.)

signals.md (next to this file) has the exact computation recipe and the claim→metric map. Read it.

Workflow

1. Pull the window. Call social_messages(asset, from_timestamp, to_timestamp). It returns {stats, messages}. Read the stats block before anything else — it carries every proportion you will cite (total_matching, unique_after_dedup, sampled, by_source, volume_curve, sentiment_balance.by_bucket, trend_words, top_channels). The sampled text only explains the stats; it never counts. Honor the strata: judge prevalence and mood ONLY from the random stratum and the stats block — head/poles are deliberately oversampled, use them for what spread and where the disagreement is. The full message set offloads to a file when large.

2. Pull the baselines (signal 1). With fetch_metric_data, get this coin's social_volume_total, social_dominance_total, a sentiment metric, and price_usd over a trailing window ≥10× the spike window (e.g. 90d for a 24h spike), at the spike's interval. If a metric name is uncertain, resolve it with metrics_and_assets_discovery. In execute, compute the spike window's percentile rank and z-score against that trailing distribution. This is signal 1.

3. Compute organic-vs-manufactured (signal 2) in execute over the messages file: organic share from copies, top-channel / top-author concentration, and the acceleration shape from volume_curve. Recipe in signals.md. Output a percentage and a one-word verdict.

4. Discover the topics, claims, and price levels. Start from the full-population trend_words — that IS the discovery: rank the dominant themes by their count/share of volume (the number carries it, the sampled text only explains each). Group raw trend words into 3–6 themes and give each one line of what's actually being said. From those themes + the head/poles text, pull the concrete, checkable claims (named actors, flows, events, targets) for signal 3 and drop vague mood. For monster windows, partition the messages file by source and spawn one research-subagent per slice (task) to extract themes + claims + verbatim quotes in parallel, then merge. In execute, extract price numbers from message text and cluster them into levels (recipe in signals.md).

5. Test narrative against chain (signal 3). For each checkable claim, pull the on-chain metric that would confirm or refute it (fetch_metric_data — exchange flows, supply on exchanges, whale txns, active addresses; map in signals.md) and label it confirmed / diverges / unverifiable with the number. For the 1–2 most consequential factual claims (a partnership, a listing, a hack), corroborate with web_search before repeating them as fact.

6. Deliver via submit_report in the format below. Then answer follow-ups from the data already pulled — re-slice the messages file or pull one more metric rather than re-running everything.

Cheap-tier note: steps 3–4 (mechanical extraction over slices) are the map step — fine for research-subagents on the utility model; keep the synthesis and the chain-divergence judgment (step 5) on your own context.

Output format (positioning readout, not an article)

Lead with the verdict. Every bullet carries its denominator/baseline.

• POSITIONING VERDICT — 1–2 lines: direction (long / short / split), conviction, and the extremeness percentile. "Crowd is aggressively long $X at a 94th-pct social spike; 68% organic, accelerating; chain confirms accumulation." This is the line they paid for — make it stand alone.
• WHAT'S DRIVING IT — the 3–6 dominant topics from the full-population trend_words, each as theme → share/count of volume → one line of what's being said. This answers "what happened"; it is the discovery layer, so it leads the body. Keep it signal-flavored — every theme carries its number; rank by share, not by how interesting the quote is. Flag any topic where the crowd splits (both sides, which has more voices) using the head/poles strata.
• EXTREME? — social volume Nth pct of trailing 90d (z=…); dominance Nth pct; sentiment skew N% bull vs M% trailing median. Spike vs baseline, always.
• ORGANIC? — % organic (unique/total), top-3 channels = N% of volume, max copies ×N, acceleration (rising / peaked / fading). Verdict: organic / mixed / manufactured.
• NARRATIVE vs CHAIN — one row per checkable claim: claim → chain metric → confirmed / diverges (the number) / unverifiable. Lead with any divergence — it's the highest-value finding here.
• CROWD PRICE LEVELS — clustered: support @ L1 (N mentions), current, resistance/target @ L2 (M mentions). Where the crowd places its line.
• WHERE IT'S HAPPENING — the "where was it discussed most" deliverable. Name only sources a reader can open: twitter accounts (twitter.com/<screen_name>) and subreddits (reddit.com/r/<sub>), plus any head-message urls. Do not print telegram/discord channel IDs (raw numeric chat_ids — unvisitable, meaningless to a reader); report those as an aggregate ("3 telegram channels ≈ 40% of volume") without the id. Internal social data has no URL — cite it as "Santiment social messages" per the citation rules.
• The denominator, always: "based on N sampled of M matching, sources: …".

Discipline (the anti-journalism rules)

• No adjective without a number. "Spiking" → "94th pct". "Lots of" → "71%". "Some say" → "N voices".
• Lead with divergence and extremeness; bury the recap. If everything is normal and organic and the chain agrees, say that in one line — a calm, confirmed read is also an answer.
• Honest denominator on every proportion; never estimate a share from the sampled text when the stats block has the full-population number.
• Analysis, not financial advice.