PM Manipulation

Read memory/MEMORY.md for context. Read the last 3 days of memory/logs/ to avoid re-flagging markets you already covered, and to compare current readings against prior ones.

Voice

If soul/SOUL.md and soul/STYLE.md are populated, match the operator's voice in the notification and write-up. If empty or absent, use a clear, direct, neutral tone. The methodology itself is the same either way.

Why this skill exists

English-language coverage of any prediction market is dense for US politics, sparse for everything else. Election-, geopolitics-, and conflict-related markets are most often pushed by actors closer to the home country than to the English-speaking financial press. By the time a coordinated narrative lands in mainstream financial outlets, the trade is over. Local-language press, regional outlets, and country-specific Telegram / X channels move first — sometimes by 24–72h.

This skill looks at the past 3 days of activity on a configurable prediction-market platform and asks: where does the price action diverge from organic English-press coverage, and does a multilingual sweep reveal a narrative push that explains it?

Configuration

Read memory/topics/prediction-markets.md if it exists for an optional ## Platform line naming the target API root. Defaults to Polymarket's public Gamma + CLOB APIs (gamma-api.polymarket.com, clob.polymarket.com, data-api.polymarket.com). The candidate selection, scoring rubric, and multilingual sweep are platform-agnostic; only the endpoints differ.

The keyword filter and locale table below are starting points — the operator can edit memory/topics/prediction-markets.md to add ## Keywords and ## Locales sections that override the defaults.

Network note

curl works — there is no network sandbox. For every curl call, if it fails or returns empty, use WebFetch as a fallback for a flaky public GET. The Polymarket APIs above are public (no auth). For news searches, prefer WebSearch with locale-specific queries — a built-in Claude tool that needs no curl.

Steps

1. Build the candidate set (markets active in past 3 days)

Pull the most-active markets across the categories most prone to manipulation: politics, geopolitics, conflict, elections (national and regional), regulatory rulings, and resolution-disputed markets.

# Top markets by 24h volume — repeat for last 3 days using volume24hr / volume1wk fields
curl -s "https://gamma-api.polymarket.com/markets?closed=false&order=volume24hr&ascending=false&limit=30"

# Also fetch by 7d volume to catch slower-burn manipulation
curl -s "https://gamma-api.polymarket.com/markets?closed=false&order=volume1wk&ascending=false&limit=30"

From the union of those two lists, pick 6–10 candidates that meet at least one of:

• Question references a non-US country, region, or conflict
• Question references a regulatory body, election, or coup/conflict event
• 7d volume > $500k AND 24h volume > $50k (real money, not just liquidity-mining)
• Slug contains keywords (default set; operator can override in memory/topics/prediction-markets.md): russia, iran, israel, china, taiwan, ukraine, venezuela, argentina, mexico, brazil, india, pakistan, nigeria, france, germany, italy, spain, eu-, nato, cartel, coup, nuclear, assassination, ceasefire, election, vote, referendum, oracle, dispute, umip

Skip anything sports, weather, or pure crypto-price (BTC > $X by Y) — those have different manipulation signatures and aren't this skill's job.

2. Pull 3-day price + trade data per candidate

For each candidate market, get the YES-token (clobTokenIds[0]) and pull 3-day price history at hourly fidelity:

curl -s "https://clob.polymarket.com/prices-history?market=$TOKEN_ID&interval=1w&fidelity=60"

Compute, from the past 72h window only:

• 3d open / close / high / low (price 0.0–1.0)
• Largest single-hour move (absolute pp change)
• Move concentration — what % of the 3d net move happened in the single largest 6h block? If > 70%, flag concentrated-move.
• Volume spike ratio — (peak hour volume) / (median hour volume). If > 8x, flag volume-spike.
• Direction inversion — did price reverse > 5pp within the 72h window? Flag reversal — common when a manipulator unwinds.

Also pull the trades feed if available (some markets only) for whale concentration:

curl -s "https://data-api.polymarket.com/trades?market=$TOKEN_ID&limit=500"

Compute whale share — % of 3d volume from the top-3 wallet addresses. If > 40%, flag whale-concentrated.

If the trades endpoint doesn't return per-market data, skip whale-share for that candidate — note it in the report rather than failing.

3. Pull comments and detect coordination patterns

For each candidate, fetch comments from the past 3 days:

# Get the event id first (comments live on Event, not Market)
curl -s "https://gamma-api.polymarket.com/events?slug=$EVENT_SLUG&limit=1"

# Top comments by reactions
curl -s "https://gamma-api.polymarket.com/comments?parent_entity_type=Event&parent_entity_id=$EVENT_ID&limit=50&order=reactionCount&ascending=false"

# Most recent
curl -s "https://gamma-api.polymarket.com/comments?parent_entity_type=Event&parent_entity_id=$EVENT_ID&limit=50&order=createdAt&ascending=false"

Filter to comments from the last 72h (createdAt within window). Then look for:

• Burst posting — > 10 comments from accounts with usernames matching the same regex pattern (e.g. user\d{4,}, common spam-bot signature) within a 6h window. Flag bot-burst.
• Narrative concentration — > 40% of top-reacted comments push the same one-sided talking point with near-identical phrasing. Flag narrative-push.
• Cross-market spam — same usernames repeating identical text across multiple unrelated markets (compare against any logged comments from prior pm-manipulation runs). Flag cross-market-spam.

Apply a reactionCount > 1 pre-filter before counting "real" reactions, but do count low-reaction comments toward bot-burst and narrative-concentration signals — that's exactly where the spam shows up.

4. Multilingual press sweep — the actual differentiator

For each candidate that hit any anomaly flag in steps 2 or 3, run a localized press sweep targeting the country/region most relevant to the question. The goal is to find coverage that explains (or doesn't explain) the price action, in the language of the place where the news would actually break first.

Pick 2–3 languages per market based on the topic. Suggested locale targets:

For each market × language pair:

WebSearch: <market keywords in target language> <site filters from table> [past 7 days]

Translate keywords into the target language before searching. If you're not confident in a translation, ask via WebFetch a translation tool URL or skip that language with a note.

Also check non-press signals where coverage tends to break first:

• Telegram channels (use WebSearch t.me/<channel>-style queries, plus general queries like "<topic in language>" telegram)
• Local-language X/Twitter via WebSearch "<topic in language>" twitter.com
• Country-specific Reddit (site:reddit.com/r/<country>)

For each candidate market record:

• Local-press timeline — when did the first regional-language story drop? Was it before, during, or after the price move?
• Narrative direction — does local press support the price move's direction, contradict it, or stay silent?
• Coverage asymmetry — is the story dominant in one language and absent in English? That's the strongest manipulation tell — local actors trading their information advantage.

5. Score each candidate (0–5 manipulation suspicion)

Sum the flags. Each is worth 1 point:

• concentrated-move (>70% of 3d move in 6h)
• volume-spike (peak/median > 8x)
• whale-concentrated (top-3 wallets > 40% volume)
• bot-burst OR narrative-push OR cross-market-spam (any one of these — comments only count once)
• reversal (>5pp price reversal within 72h)
• Coverage asymmetry — local press strongly supports OR contradicts the move while English press is silent (worth 1 point; if asymmetry is strong AND timing matches the price move within ±12h, worth 2 points and cap the total at 5)

Classify:

• 0–1 — clean. No write-up, just log the slug and stop.
• 2 — watch. Brief note in the report, no notification.
• 3 — suspicious. Full write-up + notification.
• 4–5 — high-confidence manipulation pattern. Full write-up + urgent notification, file an issue under memory/issues/ with severity medium, category unknown, detected_by pm-manipulation.

Be honest about uncertainty. "Suspicious" is not "proof". The skill's value is putting a watchlist in front of a human, not adjudicating fraud.

6. Format the briefing

Build the body in a temp file (multi-line content; never argv-pipe long strings):

TEMP=$(mktemp -t pm-manipulation.XXXXXX.md)
cat > "$TEMP" <<'MSG'
PM Manipulation Watch — ${today} (past 3d)

scanned: N markets · flagged: M · suspicious: K · high-conf: J

--- SUSPICIOUS / HIGH-CONFIDENCE ---

1. "[market question]"  — score X/5
   slug: <event-slug>
   3d: $opens → $closes ($change pp), peak vol $X
   flags: <comma-separated flag names>
   local press: <one-line summary of what foreign-language coverage said and when>
   asymmetry: <english-silent | english-contradicts | english-confirms | n/a>
   take: <one-sentence opinion — coordinated push? insider? unclear?>

2. ...

--- WATCH (score 2) ---
- "[question]" — <one-line note>

--- CLEAN (scanned but no flags) ---
N markets, no anomalies above threshold.

read it: output/articles/pm-manipulation-${today}.md
MSG

Keep the notification under 3500 chars. If it exceeds, drop the CLEAN section and the WATCH bullets first; never truncate the SUSPICIOUS section mid-entry.

7. Save the full report

Write the unabridged report to output/articles/pm-manipulation-${today}.md:

# PM Manipulation Watch — ${today}

**Window:** past 3 days · **Scanned:** N markets · **Suspicious (≥3):** K · **High-confidence (≥4):** J

## Methodology
3-day price/volume/comment scan + multilingual press sweep. Each candidate scored 0–5 across 6 anomaly classes. See `skills/pm-manipulation/SKILL.md` for the full scoring rubric.

## Suspicious Markets

### 1. [Market question] — score X/5
- **Slug:** <event-slug>
- **3d action:** $open → $close (change pp), peak hour vol $X, whale share Y%
- **Flags triggered:** <list>
- **Comments:** <coordinated patterns observed>
- **Local press:**
  - <language>: <outlet> — <one-line summary, date, link>
  - <language>: <outlet> — ...
- **English press:** <outlet — date — link, or "silent">
- **Asymmetry:** <english-silent | english-contradicts | english-confirms | n/a>
- **Take:** <2–3 sentences>

[repeat per suspicious market]

## Watchlist (score 2)
- ...

## Clean
[brief table: market, 3d change, why scanned]

Cite every source with a URL. No untranslated quotes longer than ~10 words — render the original AND a short English gloss.

8. Notify

./notify -f "$TEMP"

If status is high-confidence (≥4), prepend [URGENT] to the notification subject so it surfaces in chat.

9. Log

Append to memory/logs/${today}.md:

## PM Manipulation
- **Scanned:** N markets (past 3d)
- **Suspicious (≥3):** K
- **High-confidence (≥4):** J
- **Top flag:** "[market]" — score X/5
- **Languages searched:** <list>
- **Issues filed:** [ISS-NNN list or "none"]
- **Notification sent:** yes / no
- PM_MANIPULATION_OK

Guidelines

• Past 3 days only. Older windows belong to one-off audits. Recency matters — manipulation is a near-term phenomenon, and stale signal turns this skill into noise.
• Multilingual is the differentiator. A pure English scan is just a worse monitor-polymarket. The signal lives in coverage gaps between languages.
• No accusations. Use words like "suspected", "consistent with", "anomalous", "asymmetric coverage". Never name actors as manipulators without direct evidence.
• Translate inline. When quoting non-English press, give the original phrase + English gloss. Never paraphrase a foreign source as if it were English.
• No double-counting. If a market was flagged in the last 3 days of pm-manipulation logs and the situation hasn't escalated, mention it as "ongoing" rather than re-running the full write-up.
• Quiet weeks are useful. "Scanned 8 markets, all clean" is a valid output — flagging this skill as silent has signal of its own.

Environment Variables Required

• None — uses public Polymarket APIs + WebSearch/WebFetch
• Notification channels configured via repo secrets (see CLAUDE.md)