Scientific Hypothesis Generation

Hypothesis Generation is one of the skills bundled with K-Dense's Claude Scientific Writer (MIT). It helps a researcher turn an observation into a transparent set of candidate explanations and the tests that could tell them apart. Its core rule is that a hypothesis is a proposal to be challenged, never a finding. Version 2.1 of the skill was last reviewed by K-Dense on 2026-07-23.

Key Features

• Separate objects: the skill keeps observation, research question, hypothesis, mechanism, causal estimand, prediction, null hypothesis, negative control, and evidence as distinct labels so a mechanistic story is never passed off as a prediction.
• 12-step workflow: scope and safety gate, freezing the observation, framing the question (PICO/PICOT, PECO, diagnostic, or prognostic frames), a dated evidence boundary, rival generation, claim type and estimand, discriminating predictions, operationalization, design and analysis, anti-HARKing preregistration, replication, and human accountability.
• Rivals first: it asks for alternative explanations such as measurement artifacts, confounding, selection, collider bias, and reverse causation before any test is chosen.
• Local CLIs: seven dependency-free Python 3.11 scripts validate a hypothesis record, a prediction/rival matrix, falsification controls, an evidence ledger, and causal-claim wording, and can scaffold a preregistration. They run offline, need no API key, and deliberately do not score or rank hypotheses.

Use Cases

• Turning a surprising pilot result into testable candidates before writing a grant or protocol.
• Drafting a preregistration with outcomes, exclusions, and multiplicity decided in advance.
• Checking a draft for causal language that the study design cannot support.

Pricing and Access

The skill is free and open source under MIT. The CLIs make no network calls, so the only cost is the model session you run it in, for example Claude Code.

Getting Started

1. In Claude Code, run /plugin marketplace add https://github.com/K-Dense-AI/claude-scientific-writer, then /plugin install claude-scientific-writer.
2. Run /claude-scientific-writer:scientific-writer-init in your project.
3. Ask for a hypothesis plan, then validate the files locally, for example python3 scripts/validate_prediction_matrix.py matrix.csv.

Limitation: the skill refuses to present a hypothesis as evidence, to claim novelty from a quick search, or to give clinical or dual-use operational detail. Script output checks structure and consistency only; it does not verify scientific truth.

Frequently Asked Questions

Does it pick the best hypothesis for me?

No. The tools are explicitly non-scoring. Selection stays with the accountable human.

Do I need an API key?

Not for the bundled scripts. Literature searches in the wider workflow use other skills such as Research Lookup.

Alternatives

• Scientific Critical Thinking Skill: evaluates reasoning and evidence quality rather than generating candidates.
• Literature Review Skill: builds the evidence base the hypotheses should rest on.
• Scholar Evaluation Skill: assesses finished scholarly work.

Conclusion

Use Hypothesis Generation when you want rival explanations, falsifiable predictions, and a preregistration draft with clear human sign-off. It is a structure and discipline tool, not an oracle. Browse more research tools in the skills hub.