How to use the Scientific Hypothesis Generation agent skill

What the hypothesis-generation skill makes the agent do

An agent skill is a folder of instructions and reference files that an AI agent loads when a task matches the skill's description. This one is for scientific reasoning: it helps you move from 'we saw this pattern' to a testable plan without treating any hypothesis as a fact.

The skill keeps several objects distinct, including the observation, the research question, the hypothesis, the mechanism, the prediction, the alternative explanation, and the evidence. It generates candidate hypotheses from different explanatory classes, such as measurement artifacts, confounding, selection, and reverse causation, and labels every candidate as a candidate.

• Freezes the observation with its source, population, uncertainty and exclusions before interpreting it.
• Frames a research question, using PICO or other frameworks only where they fit.
• Declares whether the claim is descriptive, associational, predictive, causal or mechanistic.
• Derives predictions that would discriminate between rival explanations.
• Uses bundled templates for a search boundary, an evidence ledger and a preregistration scaffold, plus reference notes on causal claims and experimental design.

Install the hypothesis-generation skill from a ZIP file

Download the ZIP for the skill from the directory page. Unzip it so that the skill folder, which contains SKILL.md, sits inside one of the skills folders below. Use your system's unzip tool or file manager.

You can install it for yourself across all projects, or for a single project. Claude apps users can also upload the ZIP through the settings menu instead of copying files.

• Personal install: place the folder inside ~/.claude/skills/.
• Project install: place the folder inside .claude/skills/ in your project directory.
• Claude apps: go to Settings, then Capabilities, then Skills, and upload the ZIP.

Trigger the skill with natural requests

You do not need a special command. Describe your situation in plain language that matches the skill's purpose: turning observations or preliminary findings into a testable research plan. The skill is meant for requests about rival explanations, hypotheses, predictions, causal claims or preregistration-ready analysis plans.

Include the observation and the claim you care about. The more specific you are about the population, measurement and time frame, the more useful the output will be.

• Help me list rival explanations for the pattern in these patient-free survey results and say which tests would separate them.
• Turn this preliminary finding into a research question, candidate hypotheses and a prediction table.
• Does my data support a causal claim or only an association? Walk me through the confounders.
• Draft a preregistration-ready analysis plan for this observation, marked as an unregistered draft.

Example: from an observation to a testable plan

Suppose you notice that students who use a study app score higher on exams. The skill would first record that observation as reported, not causal. It would then separate the research question from the claim, list candidates such as the app improving study habits, students who are already motivated choosing the app, and grade inflation in one class.

For each candidate, it would state a prediction that the candidates disagree on, such as what should happen to scores in a cohort that could not self-select. It would then flag what remains an assumption and suggest a preregistration draft for the analysis.

What the skill needs and its safety limits

The listing shows no required tools and states that the skill runs no scripts and uses no network. The SKILL.md compatibility note says the bundled command-line tools are deterministic, local-only and need no credentials, models or external packages, and it targets Python 3.11+ standard library. Check the files you download before relying on this.

The skill has explicit safety gates. Before using unpublished, sensitive, controlled, personal, proprietary or security-relevant material, you must confirm authorization and applicable policies, and keep the material local unless an authorized person approves an external destination. It refuses clinical advice on specific patients, and for pathogens, toxins, weapons or similar misuse it stops at a high-level oversight note and routes you to qualified oversight.

• Do not paste sensitive or unpublished data into web searches or external AI tools without authorization.
• No script or template approval counts as an ethics, safety, regulatory or scientific approval.
• The skill does not run literature searches and does not score, rank, accept or reject hypotheses.

Tips for better results

Give the skill the raw observation before your interpretation, and say whether the pattern was expected or noticed after looking at results. The skill treats that distinction as important, and it changes which claims are allowed.

Run your own literature search and bring the results in. The skill's search-boundary template records what was searched, which is different from proving that no prior work exists.

• State your intended use and who owns the decision.
• Ask for the rival explanations before asking for tests.
• Ask whether a claim is causal, associational or predictive; the answer changes the required design.
• Keep the generated preregistration as an unregistered draft until your institution has reviewed it.