Stats
Actions
Tags
From sciagent-skills
Turns observations into testable hypotheses with predictions, mechanisms, and experiments. Follows scientific method; use for ideation or LLM testing on datasets.
How this skill is triggered — by the user, by Claude, or both
Slash command
/sciagent-skills:hypothesis-generationThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Hypothesis generation is a systematic process for developing testable mechanistic explanations from observations. This knowhow covers the full cycle: from understanding a phenomenon through literature synthesis, generating competing hypotheses, evaluating hypothesis quality, designing experimental tests, and formulating testable predictions.
Hypothesis generation is a systematic process for developing testable mechanistic explanations from observations. This knowhow covers the full cycle: from understanding a phenomenon through literature synthesis, generating competing hypotheses, evaluating hypothesis quality, designing experimental tests, and formulating testable predictions.
Good hypotheses are mechanistic (explain HOW/WHY), not descriptive (restate WHAT).
| Criterion | Definition | Example of Strong | Example of Weak |
|---|---|---|---|
| Testability | Can be empirically investigated | "Protein X binds to receptor Y" (can test with co-IP) | "Life force drives cellular growth" (untestable) |
| Falsifiability | Specific observations would disprove it | "If X is absent, effect disappears" | "X contributes to the effect somehow" |
| Parsimony | Simplest explanation fitting the evidence | Single mechanism | Multi-step chain without evidence |
| Explanatory Power | Accounts for observed patterns | Explains dose-response and tissue specificity | Explains only one observation |
| Scope | Range of phenomena covered | Applies across related systems | Limited to single dataset |
| Consistency | Aligns with established knowledge | Consistent with known pathway biology | Contradicts thermodynamics |
| Novelty | Offers new insight | Proposes unexplored mechanism | Restates established knowledge |
Hypotheses can operate at different scales. Strong hypothesis sets include explanations at multiple levels:
What is your starting point?
├── Specific observation / data → Follow the full 8-step Workflow below
├── Broad research question → Start with Step 2 (literature search) to narrow scope
├── Existing hypothesis to refine → Start at Step 5 (evaluate quality) and iterate
└── Need creative ideation first → Use scientific-brainstorming skill, then return here
| Starting Situation | Approach | Key Steps |
|---|---|---|
| Unexpected experimental result | Phenomenon-driven | Steps 1→2→3→4 (focus on competing explanations) |
| Literature gap identified | Gap-driven | Steps 2→3→4→5 (focus on novelty criterion) |
| Cross-domain analogy noticed | Analogy-driven | Steps 1→4→5→6 (focus on translating mechanism) |
| Contradictory findings in literature | Conflict-driven | Steps 2→3→4→7 (focus on discriminating predictions) |
| Large dataset patterns | Data-driven | Use hypogenic first, then Steps 5→6→7 here |
Always generate competing hypotheses (3–5): A single hypothesis is a confirmation trap. Multiple competing explanations force you to design experiments that discriminate between alternatives, not just confirm your favorite.
Start with mechanism, not correlation: "X is associated with Y" is not a hypothesis. "X causes Y via mechanism Z" is. Always include the mechanistic link (HOW the cause produces the effect).
Make predictions that differ between hypotheses: The most valuable predictions are those where Hypothesis A predicts outcome X and Hypothesis B predicts outcome Y. This is called a "crucial experiment" — design your tests around these discriminating predictions.
Ground every hypothesis in evidence: Cite existing literature for each hypothesis. "It is known that pathway X can regulate process Y [Author, 2023]; therefore, we hypothesize that..." Unsupported hypotheses are speculation, not science.
State falsification criteria explicitly: For each hypothesis, write "This hypothesis would be falsified if..." before designing experiments. If you cannot state falsification criteria, the hypothesis is untestable.
Consider the null hypothesis: The simplest explanation — that there is no novel mechanism and observed effects are due to known processes, artifact, or chance — should always be included as one of the competing hypotheses.
Scale predictions quantitatively when possible: "Expression should increase" is weaker than "Expression should increase 2–5 fold (based on known pathway kinetics)." Quantitative predictions enable power analysis for experimental design.
Confirmation bias in hypothesis selection: Generating one "main" hypothesis and 2-3 weak alternatives to make the main one look good. How to avoid: Generate hypotheses independently, then rank them by quality criteria. Have someone else review whether alternatives are genuinely competitive.
Untestable "just-so" stories: Hypotheses that sound plausible but cannot be empirically tested with current technology. How to avoid: For each hypothesis, immediately write the experiment that would test it. If you cannot design an experiment, the hypothesis needs revision.
Confusing correlation-based claims with mechanistic hypotheses: "Gene X is upregulated in disease Y" is not a hypothesis. How to avoid: Always include HOW and WHY in the hypothesis statement. Use the template: "[Mechanism] leads to [effect] because [rationale]."
Ignoring contradictory evidence: Cherry-picking literature that supports your hypothesis while ignoring opposing data. How to avoid: In Step 3 (Synthesize Evidence), explicitly section contradictory findings. Each hypothesis must address how it handles conflicting data.
Scope creep in hypothesis evaluation: Trying to make one hypothesis explain everything. How to avoid: A hypothesis does not need to explain all observations — it needs to explain the specific phenomenon under investigation. State scope boundaries explicitly.
Designing experiments that can only confirm: If your experiment cannot produce a negative result, it does not test your hypothesis. How to avoid: For each experiment, write down what "failure" looks like. Include negative and positive controls.
Neglecting feasibility in experimental design: Proposing experiments requiring technology, samples, or timelines that are unrealistic. How to avoid: Include feasibility assessment (available reagents, equipment, sample access, timeline) alongside each experimental proposal.
Understand the phenomenon: Clarify the core observation, define scope and boundaries, note what is known vs uncertain, identify the relevant scientific domain(s)
Conduct literature search: Search PubMed (biomedical) and general databases for reviews, primary research, related mechanisms, and analogous systems. Look for gaps, contradictions, and unresolved debates
Synthesize existing evidence: Summarize current understanding, identify established mechanisms that may apply, note conflicting evidence, recognize knowledge gaps, find cross-domain analogies
Generate 3–5 competing hypotheses: Each must be mechanistic (explain HOW/WHY), distinguishable from others, evidence-grounded, and consider different levels of explanation (molecular → population)
Evaluate hypothesis quality: Score each hypothesis against the 7 quality criteria (testability, falsifiability, parsimony, explanatory power, scope, consistency, novelty). Note strengths and weaknesses explicitly
Design experimental tests: For each viable hypothesis, propose specific experiments with: measurements, controls, methods, sample sizes, statistical approaches, and potential confounds
Formulate testable predictions: State what should be observed if the hypothesis is correct, specify expected direction and magnitude, identify conditions where predictions hold, distinguish predictions between competing hypotheses
Present structured output: Organize findings into: executive summary, competing hypotheses with evidence, testable predictions, critical comparisons, and detailed appendices (literature review, experimental protocols, quality assessments)
npx claudepluginhub jaechang-hits/sciagent-skills --plugin sciagent-skillsFormulates testable hypotheses from observations using the scientific method. Helps design experiments, propose mechanisms, and generate predictions.
Formulates structured, testable hypotheses from experimental observations using a scientific-method framework — derives predictions, proposes mechanisms, and designs experiments to test them.
Generates falsifiable, testable research hypotheses from notes and documents. Use when brainstorming hypotheses, generating research questions, or identifying testable predictions. Not for general Q&A.