hypogenic
Automated LLM-driven hypothesis generation and testing for tabular datasets; use when you need systematic exploration of empirical patterns (e.g., fraud detection, content analysis) and want to combine literature insights with data-driven hypothesis evaluation.
Best use case
hypogenic is best used when you need a repeatable AI agent workflow instead of a one-off prompt.
Automated LLM-driven hypothesis generation and testing for tabular datasets; use when you need systematic exploration of empirical patterns (e.g., fraud detection, content analysis) and want to combine literature insights with data-driven hypothesis evaluation.
Teams using hypogenic should expect a more consistent output, faster repeated execution, less prompt rewriting.
When to use this skill
- You want a reusable workflow that can be run more than once with consistent structure.
When not to use this skill
- You only need a quick one-off answer and do not need a reusable workflow.
- You cannot install or maintain the underlying files, dependencies, or repository context.
Installation
Claude Code / Cursor / Codex
Manual Installation
- Download SKILL.md from GitHub
- Place it in
.claude/skills/hypogenic/SKILL.mdinside your project - Restart your AI agent — it will auto-discover the skill
How hypogenic Compares
| Feature / Agent | hypogenic | Standard Approach |
|---|---|---|
| Platform Support | Not specified | Limited / Varies |
| Context Awareness | High | Baseline |
| Installation Complexity | Unknown | N/A |
Frequently Asked Questions
What does this skill do?
Automated LLM-driven hypothesis generation and testing for tabular datasets; use when you need systematic exploration of empirical patterns (e.g., fraud detection, content analysis) and want to combine literature insights with data-driven hypothesis evaluation.
Where can I find the source code?
You can find the source code on GitHub using the link provided at the top of the page.
SKILL.md Source
> **Source**: [https://github.com/aipoch/medical-research-skills](https://github.com/aipoch/medical-research-skills)
## When to Use
- **Exploratory analysis on a new dataset** where you want the model to propose multiple *testable* hypotheses from observed patterns (e.g., AI-generated text detection).
- **Benchmarking competing explanations** by generating a hypothesis bank and evaluating them consistently on validation/test splits.
- **Literature-informed research** where you want to extract claims from papers and refine them against real data (e.g., deception cues in reviews).
- **High-coverage hypothesis discovery** when you need both theory-driven and data-driven hypotheses, then merge/deduplicate them (Union workflows).
- **Hypothesis-driven classification/regression pipelines** for domains like fraud detection, content moderation, mental health indicators, or other empirical studies using tabular/JSON datasets.
## Key Features
- **Automated hypothesis generation (HypoGeniC)**: iteratively proposes and improves hypotheses using dataset feedback.
- **Literature + data integration (HypoRefine)**: extracts literature insights from PDFs and refines hypotheses jointly with empirical signals.
- **Union method**: mechanically merges literature-only hypotheses with HypoGeniC/HypoRefine outputs to maximize coverage and reduce redundancy.
- **Config-driven prompting**: YAML templates with variable injection (e.g., `${text_features_1}`, `${num_hypotheses}`) for generation and inference.
- **Scalable experimentation**: optional Redis caching, parallelism, and adaptive selection focusing on hard examples.
## Dependencies
- `hypogenic` (install via PyPI; version depends on your environment)
- Optional (recommended for cost/performance):
- `redis` (server; used for caching repeated LLM calls)
- Optional (required for literature/PDF workflows such as HypoRefine):
- `GROBID` (service; used for PDF preprocessing)
- `s2orc-doc2json` (PDF-to-structured conversion used in literature pipelines)
Install:
```bash
uv pip install hypogenic
```
## Example Usage
The following example is a minimal end-to-end workflow (dataset + config + CLI + Python). Adjust paths and prompts for your task.
### 1) Prepare a dataset (HuggingFace-style JSON)
Create three files:
- `./data/my_task_train.json`
- `./data/my_task_val.json`
- `./data/my_task_test.json`
Example schema (feature keys can be renamed, but must match your config placeholders):
```json
{
"text_features_1": ["Text A1", "Text A2"],
"text_features_2": ["Text B1", "Text B2"],
"label": ["Class1", "Class2"]
}
```
### 2) Create `./data/my_task/config.yaml`
```yaml
task_name: my_task
train_data_path: ./data/my_task_train.json
val_data_path: ./data/my_task_val.json
test_data_path: ./data/my_task_test.json
prompt_templates:
observations: |
Feature 1: ${text_features_1}
Feature 2: ${text_features_2}
Label: ${label}
batched_generation:
system: |
You are a scientific assistant. Propose testable, falsifiable hypotheses that map features to labels.
user: |
Given examples and labels, generate ${num_hypotheses} distinct hypotheses.
Return a JSON list of hypotheses, each with a short name and a testable statement.
inference:
system: |
You are a careful classifier. Use the provided hypothesis to predict the label.
user: |
Hypothesis: ${hypothesis}
Feature 1: ${text_features_1}
Feature 2: ${text_features_2}
Output the final answer as: "final answer: <LABEL>"
```
### 3) Run generation + inference (CLI)
```bash
# Generate hypotheses (HypoGeniC)
hypogenic_generation \
--config ./data/my_task/config.yaml \
--method hypogenic \
--num_hypotheses 20
# Evaluate generated hypotheses
hypogenic_inference \
--config ./data/my_task/config.yaml \
--hypotheses ./output/hypotheses.json
```
### 4) Run the same workflow (Python API)
```python
from hypogenic import BaseTask
import re
def extract_label(llm_output: str) -> str:
m = re.search(r"final answer:\s*(.*)", llm_output, re.IGNORECASE)
return m.group(1).strip() if m else llm_output.strip()
task = BaseTask(
config_path="./data/my_task/config.yaml",
extract_label=extract_label,
)
task.generate_hypotheses(
method="hypogenic",
num_hypotheses=20,
output_path="./output/hypotheses.json",
)
results = task.inference(
hypothesis_bank="./output/hypotheses.json",
test_data="./data/my_task_test.json",
)
print(results)
```
## Implementation Details
### Methods
- **HypoGeniC (data-driven)**
- Initializes hypotheses from a subset of training data.
- Iteratively evaluates hypotheses on validation data and replaces underperforming ones.
- Often uses *hard/challenging samples* to prompt improved hypotheses.
- **HypoRefine (literature + data)**
- Preprocesses PDFs into structured text (commonly via GROBID + conversion tooling).
- Generates a literature-derived hypothesis bank and a data-derived hypothesis bank.
- Refines both banks iteratively using performance feedback and relevance checks.
- **Union**
- Produces combined banks such as:
- `Literature ∪ HypoGeniC`
- `Literature ∪ HypoRefine`
- Focuses on coverage and deduplication rather than deeper joint optimization.
### Configuration and Prompt Parameters
- **Variable injection**: prompt templates can reference dataset fields and runtime parameters:
- `${text_features_1}`, `${text_features_2}`, … (from dataset JSON)
- `${label}` (ground truth label, typically used in observation templates)
- `${num_hypotheses}` (generation-time control)
- `${hypothesis}` (inference-time hypothesis text)
- **Label parsing (`extract_label`)**:
- Accuracy depends on extracting a label string that *exactly matches* the dataset’s `label` values.
- Default patterns often look for `final answer: ...`; customize for your output format.
### Performance/Cost Controls (Optional)
- **Redis caching**: reduces repeated LLM calls during iterative generation and evaluation.
- **Parallelism**: speeds up hypothesis testing on large datasets.
- **Adaptive selection**: prioritizes difficult examples to improve hypothesis quality over iterations.Related Skills
skill-auditor
A comprehensive auditor for any agent skill — including Manus, OpenClaw/ClawHub, Claude, LobeHub, or custom SKILL.md-based skills. Use this skill whenever a user wants to evaluate, audit, review, score, or quality-check an agent skill before publishing, updating, or deploying. Covers two hard veto gates (structural redlines + research integrity redlines), static quality scoring across 25 criteria (ISO 25010 + OpenSSF + Agent), dynamic test input generation, multi-mode execution testing, multi-layer output evaluation with five specialized category rubrics (Evidence Insight / Protocol Design / Data Analysis / Academic Writing / Other), a Research Veto that applies to all four research categories, human eval viewer generation, actionable P0/P1/P2 optimization recommendations, and automatic skill improvement that outputs a polished, production-ready SKILL.md. Also use whenever a user says "audit my skill", "evaluate my skill", "improve my skill", or wants a corrected version after evaluation.
two-sample-mr-research-planner
Generates complete two-sample Mendelian randomization (MR) research designs from a user-provided research direction. Use when users want to design, plan, or build a study using two-sample MR to test causal relationships. Triggers:"design a two-sample MR study", "build a publishable MR paper", "test whether this biomarker causally affects this disease", "generate Lite/Standard/Advanced MR plans", "screen multiple exposures with MR", "bidirectional MR design", "causal inference using GWAS summary statistics", or "I want to study X and Y using MR". Always outputs four workload configurations (Lite / Standard / Advanced / Publication+) with a recommended primary plan, step-by-step workflow, figure plan, validation strategy, minimal executable version, and publication upgrade path.
research-proposal-generator
Generates a comprehensive research proposal design based on input literature, including hypothesis, mechanism verification, and budget. Use when the user wants to design a research project from a paper.
research-grants
Write competitive research proposals for NSF, NIH, DOE, DARPA, and Taiwan's NSTC when you need agency-compliant narratives, budgets, and review-criteria alignment for a specific solicitation/FOA/BAA.
protocol-standardization
Standardize fragmented experimental steps into reproducible protocol documents when you need method organization, lab SOP drafting, or cross-operator reproducibility; missing parameters must be explicitly marked as "To be supplemented/Not provided".
prospero-registration-helper
Assists researchers in generating PROSPERO registration content for meta-analyses from a title and optional protocol. Use when the user wants to draft a PROSPERO registration form.
non-tumor-ml-research-planner
Generates complete non-tumor biomedical machine learning research designs from a user-provided research direction. Always use this skill when users want to plan bioinformatics + ML papers for non-cancer diseases (metabolic, cardiovascular, kidney, inflammatory, autoimmune, infectious, neurological, endocrine, wound healing, chronic multifactor), design diagnostic biomarker studies, combine GEO datasets with feature selection and ML modeling, or generate Lite/Standard/Advanced/Publication+ workload plans. Trigger for:"non-tumor ML study", "bioinformatics paper outside oncology", "key genes and diagnostic model for a disease", "pyroptosis/ferroptosis/senescence/autophagy + disease", "GEO datasets + machine learning", "RF + LASSO diagnostic model", "DEG + feature selection + validation", "immune infiltration + biomarker", "non-cancer biomarker paper". Trigger even for casual phrasings like "I want to study X using machine learning", "help me design a non-tumor bioinformatics paper", or "how do I build a diagnostic model for disease Y".
network-tox-docking-research-planner
Generates complete network toxicology + molecular docking research designs from a user-provided toxicant and disease/phenotype. Always use this skill when users want to investigate how an environmental toxicant, endocrine disruptor, heavy metal, food contaminant, pharmaceutical residue, or consumer product chemical may contribute to a disease through shared molecular targets, hub genes, pathways, and docking evidence. Trigger for:"network toxicology study", "toxicology mechanism paper", "target prediction + PPI + docking", "environmental pollutant and disease mechanism", "hub genes and docking for toxicant", "Lite/Standard/Advanced toxicology plan", "CTD + SwissTargetPrediction + GeneCards + STRING", "CB-Dock2 docking study", "triclosan/BPA/cadmium/PFAS + disease". Also triggers for Chinese phrasings:"网络毒理学研究设计"、"毒物机制论文"、"靶点预测+PPI+对接"、"环境污染物与疾病机制". Trigger even for casual phrasings like "I want to study how chemical X affects disease Y" or "help me design a toxicology paper". Always output four workload configurations (Lite / Standard / Advanced / Publication+) with a recommended primary plan, step-by-step workflow, figure plan, validation strategy, minimal executable version, and publication upgrade path.
meta-protocol-writer
Generates a PROSPERO-compliant Meta-analysis protocol based on Title and PICOS. Use when the user wants to write a protocol for a systematic review or meta-analysis.
hypothesis-generation
Structured scientific hypothesis formulation from observations; use when you have experimental observations or preliminary data and need testable hypotheses with predictions, mechanisms, and validation experiments.
faers-multi-drug-soc-planner
Generates complete FAERS-based multi-drug single-SOC safety comparison research designs from a user-provided drug set, comparator, and adverse event domain. Always use this skill when users want to compare safety signals across multiple drugs using FAERS or OpenFDA data within one System Organ Class (SOC) or bounded AE domain. Trigger for:"FAERS study comparing drugs within one SOC", "publishable FAERS safety comparison paper", "compare neuropsychiatric adverse events across beta-blockers", "Lite/Standard/Advanced FAERS safety plans", "active-comparator restricted disproportionality", "adjusted ROR logistic regression FAERS", "within-class head-to-head drug comparison", "pharmacovigilance signal comparison", "single-SOC PT-level FAERS design", or any phrasing like "I want to compare drug X and drug Y for adverse events in FAERS" or "build a comparative pharmacovigilance paper". Always output four workload configurations (Lite / Standard / Advanced / Publication+) with a recommended primary plan, step-by-step workflow, figure plan, validation strategy, minimal executable version, and publication upgrade path.
dual-disease-transcriptomic-ml-planner
Generates complete dual-disease transcriptomic + machine learning research designs from a user-provided disease pair. Use when users want to identify shared DEGs, common hub genes, cross-disease biomarkers, or shared molecular mechanisms between two diseases using public GEO data. Triggers:"shared biomarker study for two diseases", "dual-disease transcriptomic ML paper", "identify common DEGs between disease A and B", "cross-disease hub gene discovery", "shared DEG + PPI + ROC design", "immune infiltration shared biomarker", or "I want to study disease X and Y together". Always outputs four workload configurations (Lite / Standard / Advanced / Publication+) with a recommended primary plan, step-by-step workflow, figure plan, validation strategy, minimal executable version, and publication upgrade path.