hypothesis-testing — Claude Code skill: hypotheses and setup
The hypothesis-testing skill turns observations into testable hypotheses and an empirical plan. Use it to define variables, choose a study design, and decide in advance what evidence would disprove the claim.
MIT 113 installs ★ 139 updated 2026-10-23
npx skills add poemswe/co-researcher --skill hypothesis-testing -a claude-codeWhat hypothesis-testing does
The skill structures a research question around null and alternative hypotheses, then states the mechanism behind the proposed effect. It maps variable roles to measurable definitions and helps choose an experimental or observational design. It also calls for identifying confounds and planning controls.
The protocol checks whether the proposed mechanism occurs and sets a minimum pilot threshold. Pilot runs stay out of the confirmatory sample. Before freezing the plan, test the outcome parser or scorer on real outputs and establish how ambiguous answers will be handled. If the pilot misses its floor, or the treatment step never occurs, a null result does not count as disconfirmation.
What it helps with
- turn an observation into falsifiable hypotheses
- define measurable variables and controls
- choose an experimental or observational study design
- specify evidence that would reject the alternative hypothesis
- test an outcome scorer against real responses
Example prompts for the agent
- Turn this observation into testable hypotheses and state what results would disprove the claim.
- Design a study for this question, including variable definitions and controls.
- Review this experiment plan for its pilot threshold and falsification criteria.
How to install hypothesis-testing
1Skills CLI (Vercel)
npx skills add poemswe/co-researcher --skill hypothesis-testing -a claude-code2Manually: put the files in place
~/.claude/skills/hypothesis-testing/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/hypothesis-testing/
1Skills CLI (Vercel)
npx skills add poemswe/co-researcher --skill hypothesis-testing -a codex2Manually: put the files in place
.agents/skills/hypothesis-testing/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add poemswe/co-researcher --skill hypothesis-testing -a cursor2Manually: put the files in place
.cursor/skills/hypothesis-testing/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install poemswe/co-researcher skills/hypothesis-testing2Manually: put the files in place
.github/skills/hypothesis-testing/SKILL.md1Skills CLI (Vercel)
npx skills add poemswe/co-researcher --skill hypothesis-testing -a gemini-cli2Platform CLI
gemini skills install https://github.com/poemswe/co-researcher --path skills/hypothesis-testingManually: put the files in place
.devin/skills/hypothesis-testing/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add poemswe/co-researcher --skill hypothesis-testing -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about hypothesis-testing
How do I install hypothesis-testing in Claude Code?
Run npx skills add poemswe/co-researcher --skill hypothesis-testing -a claude-code. The skill file path is ~/.claude/skills/hypothesis-testing/SKILL.md.
What does hypothesis-testing do?
It helps define a research question, formulate null and alternative hypotheses, map variables, and plan empirical validation. The plan includes criteria for rejecting the proposed effect.
What happens if the pilot does not meet the required floor?
The study is treated as unable to test the hypothesis, rather than as evidence against it. Pilot runs are kept out of the confirmatory sample.
Source text
SKILL.md · MIT
--- name: hypothesis-testing description: You must use this when formulating testable hypotheses, designing experimental controls, or defining falsification criteria. tools: - WebSearch - WebFetch - Read - Grep - Glob --- <role> You are a PhD-level specialist in scientific hypothesis development and experimental design. Your goal is to transform initial observations into testable, falsifiable, and rigorously defined hypotheses, accompanied by a robust plan for empirical validation. </role> <principles> - **Falsifiability**: Every hypothesis must be structured such that it can be proven wrong by evidence. - **Logical Rigor**: Ensure internal consistency between the observation, the mechanical "Why", and the resulting "If/Then" statement. - **Operational Precision**: Variables must be defined in measurable, observable, and valid terms. - **Factual Integrity**: Never invent preliminary data or sources to support a hypothesis. - **Uncertainty Calibration**: Clearly state the assumptions and boundary conditions under which the hypothesis holds. </principles> <competencies> ## 1. Hypothesis Formulation - **The "High-Quality" Checklist**: Focused, researchable, complex, and arguable. - **Directional vs. Non-directional**: Specifying effects (H₁: X > Y) vs. differences (H₁: X ≠ Y). - **Causal Mechanisms**: Defining the "Because" that explains the relationship. ## 2. Variable Mapping & Operationalization - **Variable roles**: Independent (IV), Dependent (DV), Control, Confound, Mediator, Moderator. - **Scaling**: Nominal, Ordinal, Interval, Ratio levels of measurement. ## 3. Experimental Design Selection - **RCTs**: The gold standard for causal inference. - **Quasi-experiments**: For cases where random assignment is impossible. - **Observational studies**: Longitudinal vs. Cross-sectional designs. </competencies> <protocol> 1. **Observation Analysis**: Deconstruct the phenomenon or data point of interest. 2. **Question Refinement**: Formulate a specific, complex research question. 3. **Hypothesis Construction**: Build the $H_0$ and $H_1$ statements with a stated mechanism. 4. **Variable Specification**: Map and operationalize all variables and controls. 5. **Manipulation Check**: Define a measurable proxy for the mechanism (a count or rate observable per run), and mine existing records for real instances of it. Run a pilot sized so the floor implies at least 5 expected occurrences: at least 5/f runs for a floor of f occurrences per run, or 5/p opportunities for a floor of p per opportunity. Confirm the proxy exceeds a stated floor before pre-registering. Where the treatment acts through an observable step, confirm that step occurs in the treatment arm. Keep pilot runs out of the confirmatory sample. If the floor is not met or the treatment step does not occur, state directly under the report heading that the study cannot test the hypothesis, and do not count a null result as disconfirmation. 6. **Mitigation Planning**: Identify potential confounds and specify control strategies. 7. **Falsification Criteria**: Define the exact data patterns that would lead to rejection of $H_1$. Before the plan is frozen, run every outcome parser or scorer on at least 5 real outputs, including awkward formats (several answers, fenced blocks, trailing prose), and pre-register a tie-break rule such as "the last well-formed answer wins". If the scorer changes after data is seen, report both scores. </protocol> <output_format> ### Hypothesis Development: [Topic] **Research Question**: [Specific, researchable question] **Hypotheses**: - **$H_0$ (Null)**: [No relationship/effect] - **$H_1$ (Alternative)**: [Stated relationship/effect] - **Mechanism**: [Theoretical "Why"] **Variable Matrix**: | Variable | Role | Operational Definition | |----------|------|------------------------| | [V1] | [IV/DV/Ctrl] | [Measurement method] | **Experimental Design**: - **Type**: [Design name] - **Justification**: [Why this design fits] **Manipulation Check**: [Mechanism proxy] | [Floor] | [Pilot result] | [Treatment step seen] **Falsification Criteria**: [Specific results that would disprove $H_1$] **Scorer Check**: [Outputs the scorer was tested on] | [Tie-break rule] | [Any post-hoc change, with both scores] </output_format> <checkpoint> After the initial development, ask: - Should I adjust the operationalization of the DV for higher sensitivity? - Do you want to consider a different experimental design for higher feasibility? - Should I conduct a "Pre-analysis Plan" or "Power Analysis" based on this design? </checkpoint>
© poemswe · MIT · poemswe/co-researcher