Probability Analysis Skill
Auditable, objective probability judgments on future events — politics, war, regulation, elections, launches — with concrete numbers, not vague hedging.
Pure prompt framework: no scripts, no permission requests, no network side effects, no sensitive-data handling; explicitly forbids fabricating sources and limits paywalled content to short excerpts; data flow transparent. Deducted for unverified publisher, no verifiable source attribution or rollback discussion, and static inability to confirm downstream agent compliance.
EN/ZH workflows are consistent and internally coherent with reproducible steps; but instruction text only — no skill-specific tests, thin failure-feedback design, and dependency on external web-search which the repo itself records as having silently failed (DDG CAPTCHA). Deducted for weak failure modes and unverified key paths (static cap 10).
Trigger conditions are explicit (probability, fair probability, edge), scenarios, exclusions and output format are all specified, with native Chinese version. Deducted for unspecified mainland-China reachability of mandatory web-search sources and only partial boundary definition for evidence-poor cases.
Clear bilingual layering, progressive disclosure, concrete prohibitions; MIT license at repo level. Deducted for no per-skill version, changelog, FAQ or worked examples, and no maintenance/update path stated inside the skill files themselves.
Structured conditional decomposition, evidence weighting and Bayesian update templates with a directly usable output format give marginal value over unstructured prompting. Deducted because output correctness is unverified statically, subjective weights lack external validity evidence, and the repo discloses a known search-quality defect that contaminated past outputs (static cap 7).
The skill mandates URL, date, excerpt and manifest paths per evidence item — built-in traceability. Deducted for no reproducible tests or third-party execution evidence for this skill; the repo's Brier track records target other product lines, leaving a single evidence type (static cap 5).
- The skill depends on external web-search; the repo discloses the search layer once silently returned '0 results' with no evidence — verify search provider configuration before use.
- Probability outputs embed subjective weight judgments; this is an analysis framework, not fact, and should not be the sole basis for investment or trading decisions (the parent repo involves prediction-market trading).
- The skill files carry no version or changelog; cross-check commit history when pulling updates.
- Time-sensitive questions mandate live web search; mainland-China reachability of search sources is unverified.
What does this skill do, and when should you use it?
This is one of two skills bundled in the predict-raven repository, located at skills/probability-analysis/. It instructs an AI agent to follow a seven-step workflow for judging event probabilities: clarify the Yes/No settlement definition, decompose the event into 2-5 necessary conditions, design searches and catalog evidence by source category, assign direction and weight to each item, build a structured conditional-probability model, run Bayesian-style updates, and output a single headline probability with an 80% subjective confidence interval plus evidence tables. By default it excludes prediction-market prices, making it a market-blind probability-research method. The whole flow emphasizes auditability: every piece of evidence keeps its date, source, URL, excerpt and reliability rating.
Takes a user's event-judgment request; pins down trigger conditions, deadlines, timezones and official resolution criteria; decomposes the event into 2-5 necessary conditions and designs search queries per condition; organizes evidence by category (official statements, mainstream media, party media, third-party analysis, political/military dynamics) and rates each item on a seven-point scale from strong support to strong opposition; models the event as a conditional chain like P(Yes)=P(A)×P(B|A)×P(C|A∩B); updates from a baseline probability item by item; produces one headline probability, an 80% confidence interval, a model decomposition table, key evidence table, up/down factors and watch points, plus an archive path or source manifest for review.
- A political observer wants the probability that a diplomatic deal is formally reached by a deadline, and needs 'talks ongoing' distinguished from 'a settleable agreement exists'.
- A trader wants an independent fair-probability estimate for an event before taking a prediction-market position, to find edge versus market price.
- An analyst is asked to re-evaluate a regulatory or court event under an evidence framework where every claim is traceable and auditable.
- A user explicitly wants market prices excluded to get an estimate independent of the market consensus.
- A tech/company event watcher needs to judge whether a product ships within a window and which watch points are most likely to trigger the outcome.
What are this skill's strengths and limitations?
- Forces a concrete probability with a confidence interval and bans 'possible/impossible' style vague answers.
- Fully auditable: every evidence item carries date, source, URL, excerpt and reliability rating, with an archive path for review.
- Market-blind by default, so output is an independent estimate rather than a restatement of market consensus — useful as an independent trading signal.
- Conditional decomposition requires each condition to come from the settlement definition's necessary conditions, keeping the model disciplined.
- SKILL.md is instructions only — no executable scripts; time-sensitive questions depend on the host agent's web search, without which evidence gathering cannot be completed.
- Weights and confidence intervals are subjective scores; output quality depends on the underlying model's reasoning and search ability, with no built-in calibration.
- The skill contains none of the repo's trading execution, risk-control or market-data features — you need other parts of predict-raven for those.
- There is no documented backtest or Brier-score validation for this skill's outputs (the repo's public Brier scoring covers the World Cup product, not stated to cover this skill).
How do you install this skill?
This skill is one of two in the predict-raven repo (Alchemist-X/predict-raven, MIT licensed): after git clone, it lives at skills/probability-analysis/SKILL.md. The README's setup instructions (pnpm install + pnpm build) cover the whole collection including trading features; no standalone installation steps are documented for this skill alone — place the skills/ folder into your Agent Skills-compatible client's skill directory; the exact location depends on your client and is not specified in the source.
How do you use this skill?
Trigger it with natural language, e.g. 'Estimate the probability this event happens by end of 2026', 'What's the fair probability on this market — is there edge?', or ask it to 're-analyze this event under an evidence framework'. The agent follows the seven-step workflow and outputs a headline probability, model decomposition, evidence tables and watch points. Note: time-sensitive questions require an agent with web-search capability; if you want certain data excluded (e.g. market prices), say so explicitly in the prompt.
How does this skill compare with similar options?
The repo positions itself as the first autonomous, continuously-running trading agent on Polymarket; this skill corresponds to its market-blind probability-research mode (as in the 2026 World Cup forecasting product), contrasted with approaches that reverse-engineer probabilities from market prices. The source does not name a specific independent comparable tool.