prediction（prediction 搜索）

When to Use

User needs a defended forecast about what will happen, when it will happen, or how likely it is. Agent handles question design, base-rate search, reference-class selection, inside-vs-outside view balancing, explicit probability assignment, and after-action scoring.

Use it for business, product, technical, operational, policy, sports, market, or personal planning questions whenever the task is to forecast an uncertain outcome rather than just explain the present.

Architecture

Memory lives in ~/prediction/. If ~/prediction/ does not exist, run setup.md. See memory-template.md for structure.

~/prediction/
├── memory.md             # Activation rules, forecasting defaults, and durable lessons
├── forecast-log.md       # Open forecasts with probability, horizon, and next review date
├── scorecard.md          # Resolved forecasts, Brier scores, and error patterns
├── reference-classes.md  # Reusable base-rate cases by domain
├── assumptions.md        # Active drivers, fragilities, and update triggers
└── archive/              # Old resolved periods and retired forecasting themes

Quick Reference

Use the smallest file that resolves the blocker.

Requirements

• No credentials or external services are required by default.
• Ask before storing sensitive personal forecasts, legal matters, health outcomes, or unreleased company information.
• Prefer questions with a clear resolution rule and time horizon. If those are missing, define them before assigning a probability.

Prediction Contract

Every serious forecast should leave behind:

1. the exact question being forecast
2. the resolution rule and deadline
3. the base rate or reference class used
4. the main drivers that could move the answer
5. a numeric probability or ranked scenario split
6. the trigger that would cause an update before resolution
7. a later score or post-mortem once reality is known

This is the minimum needed to improve accuracy instead of producing forgettable guesses.

Core Rules

1. Turn the Prompt Into a Resolvable Question First

• Use question-design.md before making any forecast that matters.
• If the target, threshold, deadline, or resolution source is fuzzy, the forecast is not auditable and the hit rate cannot improve.

2. Start With the Outside View Before the Story

• Pull a base rate or nearest reference class before building an inside-view narrative.
• Humans overweight unique details and underweight how often similar situations actually happen.

3. Run the BRACE Forecast Loop on Every Non-Trivial Prediction

• Use forecast-loop.md: Base rate, Resolution rule, Arguments both ways, Confidence assignment, Evaluation plan.
• A loop beats intuition because it forces evidence on both sides and leaves a trail for later scoring.

4. Express Uncertainty Numerically and Defend It

• Give a number, range, or explicit scenario split rather than words like "probably" or "maybe."
• Use calibration.md to map evidence quality, sample size, and model disagreement into probability levels.

5. Separate Signal From Narrative Heat

• Track what is actually predictive, what is merely interesting, and what is just recent or vivid.
• Strong stories with weak base rates are noise, not edge.

6. Update Only on Information That Changes the Odds

• Pre-commit to update triggers: deadline changes, threshold changes, a major driver flips, or new data changes the reference class.
• Constant micro-updating on every headline produces churn without better accuracy.

7. Score Every Meaningful Forecast and Learn From Misses

• Use scoring-and-review.md after resolution and store the result in the local scorecard.
• Unscored forecasts feel smart in the moment and teach nothing later.

Common Traps

These are the failure modes that usually destroy forecast accuracy even when the reasoning sounds smart.

Data Storage

Local state in ~/prediction/ may include:

• activation preferences and forecasting defaults
• open forecasts with probabilities, scenarios, and review dates
• resolved forecasts with scores and miss patterns
• reusable reference classes and base-rate notes
• assumptions and update triggers for active forecasting topics

Store only the smallest durable note that improves the next forecast.

Security & Privacy

Data that stays local:

• forecast logs, scorecards, assumptions, and reference-class notes in ~/prediction/

Data that leaves your machine:

• none by default unless the current environment uses approved search or browsing tools for evidence collection

This skill does NOT:

• claim certainty where evidence is weak
• guarantee accuracy or positive expected value
• place bets, trades, or automatic decisions on the user's behalf
• store credentials, account numbers, or private medical records
• modify its own skill file

Scope

This skill ONLY:

• turns ambiguous prediction prompts into auditable forecasting questions
• uses base rates, reference classes, and explicit probability assignments
• tracks forecast quality through scoring and post-mortems
• stores lightweight local notes that improve later predictions

This skill NEVER:

• pretend that every question is forecastable with confidence
• replace licensed legal, medical, or investment advice
• confuse eloquent explanation with predictive power
• skip scoring on forecasts that matter

Related Skills

Install with clawhub install <slug> if user confirms:

• analysis - structure assumptions, causal chains, and trade-offs before forecasting.
• compare - evaluate scenario branches and option differences after the forecast is framed.
• decide - turn probabilities and uncertainty into explicit decision choices.
• statistics - dig deeper into inference, distributions, and sampling logic behind the forecast.

Feedback

• If useful: clawhub star prediction
• Stay updated: clawhub sync