菜单
Systematic procedure for formulating testable hypotheses, designing tests, and updating beliefs based on evidence. Supports context-adaptive variants.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Route any input through a branching question tree to narrow down the optimal response strategy before writing. Two stages — PERCEIVE (classify input) then ACT (select response). Covers all prompt types.
Generate exhaustive guesses about user input using ALL search methods with coverage tracking. Guessing is SEARCH through possibility space. Tracks space created vs space covered to ensure comprehensive exploration.
Systematically evaluate and select from a set of guesses, options, or possibilities. Combines ARAW analysis with prioritization to determine which guesses are strong, weak, actionable, or eliminable.
Extract hidden assumptions from any content. Surfaces what must be true for claims to hold, enabling deeper analysis.
Sub-orchestrator for analytical requests. Routes to decomposition, systems analysis, comparison, risk assessment, or synthesis based on what kind of analysis is needed.
Assume Right - Deep recursive rightness search. For every claim, assume it's right — find what must follow, then assume THOSE implications are right too. Recurse until bedrock. Track every claim found.
| name | ht - Hypothesis Testing |
| description | Systematic procedure for formulating testable hypotheses, designing tests, and updating beliefs based on evidence. Supports context-adaptive variants. |
| context | fork |
| output | {"format":"prose"} |
Input: $ARGUMENTS
Before executing, identify which interpretation matches the user's input:
Interpretation 1 — Test a specific claim: The user has a concrete hypothesis or assertion and wants a rigorous framework to evaluate whether it holds up against evidence. Interpretation 2 — Validate a business assumption: The user has a belief about their market, customers, or strategy (e.g., "I think users want X") and needs to design a way to confirm or refute it. Interpretation 3 — Investigate a hunch: The user has a theory about why something is happening (e.g., "I think the bug is caused by X" or "I suspect my fatigue is from Y") and wants structured help working through it.
If ambiguous, ask: "I can help with rigorously testing a formal claim, validating a business assumption, or investigating a personal hunch — which fits?" If clear from context, proceed with the matching interpretation.
Systematic procedure for formulating testable hypotheses, designing tests, and updating beliefs based on evidence.
Default: 2x. Parse depth from $ARGUMENTS if specified (e.g., "/ht 4x [input]").
| Depth | Min Hypotheses | Min Tests per Hypothesis | Min Competing Explanations | Min Falsification Attempts |
|---|---|---|---|---|
| 1x | 2 | 1 | 1 | 1 |
| 2x | 3 | 2 | 2 | 2 |
| 4x | 5 | 3 | 3 | 3 |
| 8x | 7 | 4 | 5 | 5 |
| 16x | 10 | 6 | 7 | 8 |
These are floors. Go deeper where insight is dense. Compress where it's not.
Before full analysis, assess context:
| Factor | Value | Notes |
|---|---|---|
| Time Pressure | URGENT / NEAR / NORMAL | |
| Stakes | HIGH / MED / LOW | |
| Domain Expertise | EXPERT / INTERMEDIATE / NOVICE | |
| Test Cost | CHEAP / EXPENSIVE |
| Context | Variant | Steps |
|---|---|---|
| URGENT | HT-Lite | 1 (clarify claim), 4 (minimal test), 7 (quick conclude) |
| LOW stakes + CHEAP test | HT-Quick | 1, 4, 5, 7 |
| EXPERT + known domain | HT-Check | 2-4, 7 (skip basics, focus on test design) |
| CHEAP test + learning goal | HT-After | 1, 4, 7 + document learnings |
| HIGH stakes + EXPENSIVE test | HT-Full | All 7 steps + replication planning |
| Default | HT-Standard | All 7 steps |
Selected variant: [variant] because [reasoning]
Precisely specify what is being claimed:
STATE THE CLAIM CLEARLY
Vague: "Exercise is good for you" Clear: "30 minutes of moderate aerobic exercise 3x/week reduces risk of cardiovascular disease"
IDENTIFY THE CLAIM TYPE
Existential claims:
Universal claims:
Statistical claims:
Causal claims:
SPECIFY SCOPE CONDITIONS
IDENTIFY COMPETING CLAIMS
ASSESS BACKGROUND PLAUSIBILITY
Transform the claim into testable hypotheses:
STATE THE RESEARCH HYPOTHESIS (H1)
Good hypotheses are:
Format: "If [condition], then [prediction]"
STATE THE NULL HYPOTHESIS (H0)
Purpose: Provides a default to test against
STATE ALTERNATIVE HYPOTHESES
DERIVE SPECIFIC PREDICTIONS
From each hypothesis, derive:
More specific predictions = more informative tests
SPECIFY FALSIFICATION CRITERIA
Karl Popper's criterion: If nothing could prove it wrong, it's not scientific.
EXAMPLE:
Claim: "Mindfulness meditation reduces anxiety"
H1: Participants completing 8-week mindfulness program will show lower anxiety scores than waitlist control
H0: No difference in anxiety between groups
Alternative: Attention placebo explains any effect
Predictions:
Falsification: Would reject H1 if mindfulness group shows no improvement or performs worse than control
Estimate the prior probability of the hypothesis:
CONSIDER BASE RATES
CONSIDER THEORETICAL SUPPORT
CONSIDER PRIOR EVIDENCE
CONSIDER EXTRAORDINARY CLAIMS
ASSIGN A PRIOR PROBABILITY
Be explicit:
Priors should be:
CONSIDER SENSITIVITY TO PRIORS
CALIBRATION GUIDELINES:
Note: This is pre-experimental probability, before your test.
Design a test that could actually falsify the hypothesis:
SEVERITY PRINCIPLE
A test is severe if:
Weak tests: Would pass whether hypothesis true or false Strong tests: Would fail if hypothesis is false
INCREASE SEVERITY BY:
High-risk predictions:
Example: Weak: "Treatment will help some people" Severe: "Treatment will produce 10-point improvement in 60% of participants within 4 weeks"
Good methodology:
Multiple tests:
CHOOSE ALPHA LEVEL AND POWER
Alpha (Type I error rate):
Power (1 - Type II error rate):
SPECIFY DECISION RULE
Before seeing results, specify:
Example:
PRE-REGISTER
Commit to analysis plan before data collection:
Assess the strength of evidence from the test:
CLASSICAL HYPOTHESIS TESTING
P-value interpretation:
Standard thresholds (arbitrary conventions):
Effect size:
BAYESIAN UPDATING
Bayes' theorem: P(H|D) = P(D|H) × P(H) / P(D)
Components:
Bayes factor:
ASSESS EVIDENCE QUALITY
Was the test actually severe?
Internal validity:
External validity:
COMPARE TO PREDICTIONS
Exactly as predicted: Strong support Partially as predicted: Moderate support Opposite of predicted: Strong disconfirmation Null result: Depends on power
Note: High-powered null results are informative
Revise probability estimates based on evidence:
CALCULATE POSTERIOR PROBABILITY
Using Bayes' theorem:
Posterior odds = Prior odds × Bayes factor
Example:
Strong evidence moves beliefs substantially Weak evidence moves beliefs modestly
CONSIDER REPLICATION
Single study provides limited evidence:
Rules of thumb:
AVOID BELIEF UPDATING ERRORS
Base rate neglect:
Confirmation bias:
Anchoring:
Motivated reasoning:
DOCUMENT BELIEF CHANGE
Record:
Be willing to say:
IDENTIFY REMAINING UNCERTAINTY
What would further increase/decrease confidence? What additional evidence is needed? What are the key remaining uncertainties?
Formulate appropriate conclusions and next steps:
STATE THE CONCLUSION
Based on posterior probability:
State your conclusion at the strength the evidence supports. If the posterior is 65%, say the evidence leans toward support — don't retreat to 'inconclusive' because it feels safer.
DISTINGUISH TYPES OF CONCLUSIONS
Epistemic conclusions (about knowledge):
Practical conclusions (about action):
Evidence can be insufficient for knowledge but sufficient for practical decision
CONSIDER IMPLICATIONS
If hypothesis is supported:
If hypothesis is refuted:
If inconclusive:
SPECIFY NEXT STEPS
Additional research:
Practical actions:
Theory development:
DOCUMENT FOR FUTURE REFERENCE
Create record:
Enable cumulative knowledge building
