K-Dense-AI/science-superpowers

What it solves

It prevents common scientific pitfalls like p-hacking and HARKing (hypothesizing after results are known) by enforcing a rigorous, disciplined workflow for AI research agents. It transforms AI assistants from generalists that jump straight into data analysis into specialized co-scientists that follow a strict sequence of framing, pre-registration, and verification.

How it works

The project provides a library of composable "skills" and bootstrap instructions that automatically trigger when an agent begins a research task. Instead of immediate execution, the agent is guided through a mandatory sequence:

1. Framing: Turning a rough interest into a falsifiable question.
2. Pre-registration: Locking hypotheses and decision rules before any data is seen.
3. Reproducible Execution: Running analysis in pinned environments with fixed seeds and immutable data.
4. Verification: Root-causing anomalies and red-teaming results before reporting.

It has zero third-party dependencies and runs using only the agent harness and a POSIX shell.

Who it’s for

Researchers, scientists, and engineers using AI agents to perform computational science and data analysis who want to ensure their findings are verifiable and reproducible.

Highlights

• The Iron Law: Enforces pre-registration of predictions and decision rules to separate confirmatory from exploratory work.
• Zero-Dependency: Runs on any POSIX shell without needing external libraries.
• Automated Workflow: Skills trigger automatically via session-start hooks in supported agent harnesses (e.g., Cursor, Claude Code, Gemini CLI).
• Feasibility Mode: An opt-in path to prove a computation can run at scale before committing to the full pre-registration process.