prediction-market
Skill0 starsChanged 3 months agoMITView on GitHub
Look up what real-money prediction markets (Polymarket, Kalshi, plus Manifold and Metaculus as weaker signals) currently price a future event at, including probability, trend, volume, participants and a credibility check. Use whenever someone asks whether some public event will happen, how likely it is, what the odds are, where a stock or coin will trade, or what the market thinks — "will X happen", "what are the chances", "odds of", "how likely is", "how far will it fall", "会不会发生", "有多大可能", "概率多大", "予測市場". Read-only; it never trades. When no live market matches the question it says so rather than inventing a number.
1 other project uses a version of this.
Adopted by
Most recently taken up by
- baofeng-tech/agent-skills-sotheir own version, first seen Apr 28, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- baofeng-tech/agent-skills-so★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
The file
We split this skill into the rules it gives an agent
11 rules. About 0 are common, 9 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈9
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 11
This skill is a lookup tool, not a forecaster: If no live market covers the question, say so.
- Distinctive: line 32
Every venue lists in English, so a question in any other language finds nothing until you translate it.
- Very specific: line 83
Read rules and confirm it resolves on the thing the user actually asked about.
- Distinctive: line 88
Check the window too, not just the level: a Polymarket ladder titled "in 2026" can have opened in late 2025, so part of what it covers has …
- Distinctive: line 90
If the closest market answers a *related but different* question, say that explicitly rather than quietly substituting it.
- Distinctive: line 101
If every candidate fails this check, the answer is no market, exactly as if the search had come back empty.
- Distinctive: line 132
Put the level in one of your keyword groups ("Bitcoin 150k").
- Very specific: line 246
Run --show-dropped to see whether the market is there and being filtered.
- Distinctive: line 248
If all three come back empty, say plainly that no prediction market covers this, and stop.
- Distinctive: line 264
If they would rather not enable it, fall back to fetching these URLs with your own web-fetch tool and reading the raw JSON yourself:
- Distinctive: line 290
Place, sign, or prepare a trade.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Prediction Market Lookup
Answer "how likely is X?" with what people are actually betting, not with a guess.
This skill is a lookup tool, not a forecaster. If no live market covers the question, say so. Never present your own estimate as a market probability, and never fill a gap with a number you reasoned your way to.
Run it
python3 scripts/pm_query.py search "keyword group one" "keyword group two"
python3 scripts/pm_query.py detail polymarket <conditionId>
python3 scripts/pm_query.py detail kalshi <ticker>
python3 scripts/pm_query.py compare kalshi:<ticker> polymarket:<conditionId>
python3 scripts/pm_query.py spot META # underlying price, to anchor a ladder
python3 scripts/pm_query.py spot BTC # crypto resolves to the coin, not an ETFStdlib Python only — no install step. Paths are relative to this skill directory.
Workflow
1. Turn the question into English keywords
Every venue lists in English, so a question in any other language finds nothing until you translate it. Write two or three keyword groups covering different phrasings. Groups are scored independently, so an extra phrasing can only help.
Always include one group that is the bare subject on its own — the company, person, coin or country, with no qualifiers. Venue titles rarely contain the words a person would use ("stock", "price", "fall"), and those words are common enough across markets that they carry no signal. The bare name is what matches.
| User asks | Search |
| Will the Fed cut in September? | "Fed" "Fed rate cut September" |
| 美联储九月会降息吗 | "Fed" "Fed rate cut September" |
| Will Meta keep falling? | "Meta" "Meta Platforms" "Meta up or down" |
| Is the AI bubble going to pop? | "AI bubble" "AI bubble burst" |
| 美国明年会衰退吗 | "recession" "US recession 2027" |
2. Search
python3 scripts/pm_query.py search "Meta" "Meta Platforms" "Meta up or down"Returns live markets only, grouped under an events index that tells you how
many outcomes each event has. Settled markets are dropped, and results sharing
only a generic word with the question are filtered out.
verdict is found, no_live_market, or sources_unavailable — the last
means nothing could be queried, which is not the same as nothing existing.
Coverage is capped two ways: --limit events per venue (default 8) and
--max-outcomes per event (default 20). --limit does not widen a
truncated ladder — that needs --max-outcomes. Price ladders routinely run past
20 rungs, so on a "how far" question raise it before reading the shape. limit_note,
events_not_returned and possibly_truncated say when either bit. --show-dropped adds dropped_examples,
the markets the relevance gate rejected. Check both before concluding a market
does not exist.
Quote display_title, not title: in search output every row of an event
shares the event's title, so "What will META hit in August 2026?" plus 50.5%
does not say whether that is a rise or a fall, and "Nobel Peace Prize winner"
plus 30% names nobody. display_title folds in the row's own outcome.
3. Check the resolution criteria before believing a match
This is the step that decides whether the answer is right. Titles mislead: "Fed rate cut by December" can mean the December meeting or any cut before December, and those are different questions with different prices. A market titled "What will META hit in August" is not about the closing price — it resolves on whether any one-minute candle touches the level intraday.
Read rules and confirm it resolves on the thing the user actually asked about.
Kalshi search results carry no rules — run detail to get them. Manifold never
has any, and a market flagged "no resolution text" cannot be verified at all,
so it must not carry an answer on its own.
Check the window too, not just the level: a Polymarket ladder titled "in 2026" can have opened in late 2025, so part of what it covers has already happened. If the closest market answers a related but different question, say that explicitly rather than quietly substituting it.
A candidate surviving the search is not evidence that a market exists. The
relevance gate is deliberately a little loose so it does not hide real markets,
so a thin coincidence gets through now and then — a question about your
daughter's college application can surface a football game against the same
university, and one about your own company can surface a market whose listed
outcome merely ends in "Company". Each candidate carries matched, the words it
matched on; when that is one or two incidental words, be suspicious.
If every candidate fails this check, the answer is no market, exactly as if the search had come back empty. Do not soften it into "here is a related market's number".
4. Pull detail on the market you picked
detail adds the 30-day probability history, holder counts, order-book depth,
the full resolution text and a fetched_at timestamp. Run it before writing the
answer — search results carry no resolution text and no trend.
Comparing venues? Use compare, and only on markets you have confirmed ask the
same question. It returns spread_pp, agree, and caveats. agree is null when
agreement could not be verified — a missing end date, missing rules, or two
sides resolving on different events (confirming an IPO is not completing one).
Null is not agreement; report the spread and the caveat, not a match. The script deliberately refuses to compute a cross-venue spread
during search, because comparing "Cut 25bps" on one venue against "No change" on
another manufactures a meaningless 70-point "disagreement".
5. Price-level questions need an anchor
"How far will it fall?" is answered by a ladder — a set of rungs at
different thresholds. A rung reading 52 percent is meaningless until you know
where the price is now, so run spot TICKER first and lead with it.
Check the name the quote came back with. Bare crypto tickers resolve to
equities on the data source — BTC is an ETF trading near 28 USD, not Bitcoin
near 65,000 USD — so the script maps known coins to their pair form and says so
in note. For anything it does not know, name and instrument_type are
there to be read: a ladder about a coin anchored to something labelled ETF is
the wrong anchor.
Put the level in one of your keyword groups ("Bitcoin 150k"). Venue search is sensitive to it — the most on-point markets often surface only when the number is present — and a number in the question also protects that rung from the per-event cap.
Report the ladder as a ladder: several rungs in order, so the reader sees the distribution. Do not truncate it to three rows — the shape is the answer. Say whether the rungs are touch (any point intraday) or close levels; they price very differently.
Rungs come back grouped by event and ordered by threshold, with ↑ and ↓
sides kept apart. Three checks run over them, and each flag names the rungs it
is about:
- Ladder inconsistent — rungs within one event contradict each other. It is
stamped on the direction group it belongs to, so a clean
↓side stays clean while a broken↑side is called out. Those rungs are unquoted stubs rather than prices. - Cross-market inconsistency — the same underlying priced out of order across two events, e.g. touching 170k dearer than touching 150k.
- Window inconsistent — a shorter window priced above a longer one that contains it.
- Same level priced twice — one venue quoting the same level and deadline two different ways, e.g. a ladder rung at 2.5% beside a standalone market at 1.2% for the identical level. At least one is wrong; report the range, not one of the numbers.
- Distribution incomplete / incoherent — outcomes the venue itself
declares mutually exclusive that sum to well under 100% (some are missing)
or above it (the quotes contradict each other). Exclusivity is read from
Polymarket's
negRiskand Kalshi'smutually_exclusive, never guessed: a board of nested deadlines, a "which country gets removed" field and a playoff-berth field all sum past or under 100% while being perfectly healthy. A field the venue does not declare exclusive is left unchecked, andexclusivein the events index says which is which. - End date already passed — the market is past its date and may be awaiting settlement rather than showing a live view.
A flag makes those rungs indicative at best. Say which rungs it applies to rather than discrediting the whole answer.
These structural flags exist only in search output. They compare markets
against each other, which a single-market detail call cannot do, so detail
returns an empty or shorter flags list for the same market. Carry the search
flags forward; do not treat the detail call as the fuller picture. detail adds
credibility flags of its own, so a rung you quote may need both.
6. Write the answer
Conclusion first, no preamble. Roughly 150 words for a single market; a ladder or a two-part question will run longer, and that is correct — never drop rungs or flags to hit a word count.
7.5% — Polymarket: US enters recession before end of 2026 (as of 18 Aug, 11:42 PT)
· Trend: -1pt over the week; down from 15.5% a month ago
· Size: 1.7M USD lifetime volume, 42k USD resting, 975 wallets holding → credible
(Kalshi publishes no resting depth — quote open interest there instead)
· Resolves: NBER declares a recession, or two consecutive quarters of negative real GDP
· Cross-check: Kalshi 6%, Manifold (play money) 8.5% — all three agree
· https://polymarket.com/event/us-recession-by-end-of-2026Rules for that block:
- Never quote a price without saying which outcome it belongs to. Every
market carries
probability_ofandoutcome_prices. A market with outcomes Up and Down priced at 0.55 means Up is 55 percent — reporting a bare 55 percent on a question phrased "will it fall" states the opposite of the market's view. Check the label every time. - Write it in the user's language. The script emits English labels and flags; translate them. Never drop a ⚠️ because the answer reads better without it — every flag the script returns must appear in your answer.
- Report the probability as the market's view, never as yours. "The market prices this at 73%", not "there is a 73% chance".
- Multi-outcome events (who wins X): lead with the field's sum, not with a
name. The
eventsindex carriesoutcomes_sum. When it is far below 100% the rest of the probability sits on outcomes you were never shown, so "the favourite is X at 8.5%" is false — X is only the highest of what came back. Say how much is unaccounted for, then give the top few by probability (the script now orders them that way). When the sum is above 100% the quotes contradict each other and none of them is worth much. Ladders are different — show the rungs in order, per step 5. - A price under 2% is flagged "priced as a long shot", not settled. On a ladder tail that is ordinary, so report it as the market's view of a remote outcome.
possibly_truncatedmeans outcomes are missing — either the per-event cap bound, or the venue returned fewer than the event holds (outcomes_on_venueagainstoutcomes_returned). Raise--max-outcomes, and when the venue is the one holding back, say the field is partial rather than presenting it as complete.- Manifold is play money and Metaculus is not a market — label both, and never lead with them when a real-money venue has the same question.
- Give the market link on its own line, plain text.
7. Multi-part questions
"Will it fall, and how far?" is two questions. Answer them separately and say which part the market actually covers.
Watch for the two parts resolving on different events. A "when will X IPO" ladder can resolve on the company announcing an offering while the "will X IPO by date" market resolves on it completing one — months apart, and not comparable. Report them as separate answers; never present one as the other's cross-check. It is common for the level question to have a ladder while the direction question has only a thin daily up/down market, or for the near term to be priced and the horizon the user asked about to have no market at all. Say so plainly instead of stretching one answer over both.
When nothing matches
The script returns "verdict": "no_live_market". Before reporting that:
- Retry with the bare subject alone if you have not already.
- Try the formal name — "Meta Platforms" rather than "Meta stock", "Alphabet" rather than "Google".
- Run
--show-droppedto see whether the market is there and being filtered.
If all three come back empty, say plainly that no prediction market covers this, and stop. You may name the nearest related markets and explain how they differ. You may not produce a probability.
Most personal questions ("will I get promoted", "will my offer be accepted") have no market and never will. That is a fine answer.
Platform notes
Claude Code / Codex / Cursor — works as-is.
claude.ai — the skill sandbox blocks outbound network by default, so the script will fail with a connection error. Tell the user once:
Settings → Capabilities → Allow Network Egress → Domain allowlist → All domains
If they would rather not enable it, fall back to fetching these URLs with your own web-fetch tool and reading the raw JSON yourself:
https://gamma-api.polymarket.com/public-search?q=KEYWORD&limit_per_type=5&events_status=activehttps://api.elections.kalshi.com/v1/search/series?query=KEYWORD
The fallback is worse: responses are large, and every filter the script applies
(settled markets, stale prices, relevance, ladder integrity, credibility) you
must then apply by hand. references/sources.md lists the traps.
Metaculus is off by default. It needs the scrapling CLI on PATH (or
PREDICTION_MARKET_SCRAPLING pointing at the binary) and takes ~40s. Add it to
the full list — --sources polymarket,kalshi,manifold,metaculus — because
--sources replaces the default rather than extending it. Passing
--sources metaculus alone queries nothing else, and on a machine without
scrapling that returns sources_unavailable.
Deeper reference
references/sources.md— endpoints, field mappings, and the silent failures each venue's API is prone to. Read before changing the script or hand-fetching.references/interpreting.md— how to read a probability, when volume makes a price meaningless, how to read a ladder, and what cross-venue disagreement means.
Never
- Place, sign, or prepare a trade. This skill is read-only and holds no keys.
- Give investment advice.
- Present a settled market's final price as a current probability.
- Report a number without its credibility flags, or a price without its outcome label.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

