Science Superpowers

Science Superpowers is a complete computational-science methodology for your research agents, built on a set of composable skills plus initial instructions that make sure your agent actually uses them. It has zero third-party dependencies — it runs with only your agent harness and a POSIX shell.

⭐ If Science Superpowers helps your research, please star this repository. A star helps other scientists and engineers find the project and tells us the methodology is worth expanding.

Learn more: Introducing Science Superpowers — why we built it, the Iron Law, and the full workflow. Related essays are collected under From the blog.

Stay up to date: Follow K-Dense on X, LinkedIn, and YouTube for new skills, release announcements, and research workflow demos.

🎬 Prefer to watch first? Getting Started with Scientific Agent Skills covers how skills plug into your agent, and Skills 101 walks through writing one yourself.

It is a reimplementation of Superpowers (a software-development methodology) for a different domain: doing science with data. The architecture is the same — skills that auto-trigger via a session-start bootstrap — but the workflow is the research lifecycle, and the central discipline is pre-registration instead of test-driven development.

Contents

• How it works
• The basic workflow
• Example: what using it looks like
• What's inside
• Philosophy
• From the blog
• Installation
• Contributing
• License
• Star History

How it works

It starts the moment you fire up your agent. As soon as it sees you're trying to investigate something, it doesn't jump straight into running code on your data. Instead it steps back and helps you turn a fuzzy interest into a precise, falsifiable question.

Once the question is clear, it grounds the work in prior literature and standard methods, designs the analysis, and pre-registers the hypotheses, predictions, and decision rules before looking at the outcomes. That separation — confirmatory vs. exploratory, predictions locked before data — is what protects the work from p-hacking and HARKing (hypothesizing after results are known).

Then it executes the pre-registered plan in a reproducible workspace (pinned environment, fixed seeds, immutable raw data), investigates anomalies by root cause instead of quietly dropping inconvenient data, verifies every claim against fresh reproduced evidence, and red-teams the result before reporting it.

Because the skills trigger automatically, you don't need to do anything special. Your research agent just has Science Superpowers.

The basic workflow

1. framing-research-questions — Activates before any analysis. Turns a rough interest into a precise, falsifiable question with hypotheses, the data needed, and what would count as an answer. Saves a question document.
2. surveying-prior-work — Grounds the question and chosen methods in what's already known: standard methods, known confounds, prior effect sizes.
3. designing-the-analysis — Breaks the work into bite-sized analysis steps with exact datasets, variables, models/tests, power, and decision rules.
4. preregistering-analysis — The Iron Law. Locks hypotheses, directional predictions, and decision rules — and the confirmatory/exploratory split — before any outcome is seen.
5. setting-up-reproducible-analysis — Isolated, reproducible workspace: pinned environment, fixed seeds, immutable raw data, clean baseline.
6. subagent-driven-analysis or executing-analysis — Carries out the pre-registered plan with review checkpoints.
7. investigating-anomalous-results — Activates when results look wrong. Root-cause investigation before any adjustment.
8. verifying-results-before-claiming — Evidence before claims: re-run, check assumptions, robustness, reproduce.
9. requesting-red-team-review / receiving-critical-review — Adversarial review before you believe or report a result.
10. reporting-and-archiving-findings — Reproducibility check, then write-up/preprint/iterate/shelve/discard, then archive code + data + environment.

The agent checks for relevant skills before any task. Mandatory workflows, not suggestions.

Feasibility mode (opt-in)

The sequence above assumes the analysis can be run. When that is the open question — an unbenchmarked solver, a simulation that has never assembled at full size, a memory ceiling nobody has measured — paying for the full sequence first means writing plans and predictions for a study that may not exist.

For that case the agent will offer establishing-feasibility-first: a minimal end-to-end runner, three scaling probes against a budget whose kill criteria are proven to fire, then one exploratory campaign — deferring the survey, plan, and pre-registration until the measured answer is in.

It is opt-in and opt-out by you alone. The agent never enters the mode on its own initiative, never leaves it on its own initiative, and nothing produced inside it can be reported as confirmatory. Pre-registration is deferred, not cancelled: the exit gate hands back to the standard sequence at full strength, with the feasibility measurements as inputs to the design and the exploratory results as hypotheses rather than evidence.

Example: what using it looks like

You don't invoke anything. You just talk to your agent like normal — the skills trigger themselves.

You: I have last quarter's churn data. Does our new onboarding flow reduce churn?

A vanilla agent would load the CSV and start slicing. With Science Superpowers, the agent steps back first:

Agent: Before touching the data, let me frame this precisely (framing-research-questions). The question is whether the new onboarding flow causes lower 90-day churn — but users weren't randomly assigned to it, so cohort and seasonality are confounders. Here's the falsifiable version, the data we'd need, and what would count as an answer…

[surveys prior internal analyses and standard methods, then designs the analysis]

Agent: I'm now pre-registering the hypothesis, the directional prediction (new flow → lower churn), and the decision rule before I look at any outcome (preregistering-analysis). Everything beyond this that wasn't pre-registered will be labeled exploratory.

[runs the pre-registered plan in a pinned, seeded, reproducible workspace; an anomaly appears]

Agent: One cohort shows implausibly low churn. Rather than dropping it, I'm root-causing it (investigating-anomalous-results) — it turns out to be a logging gap, not a real effect.

[verifies every claim against freshly reproduced output, then red-teams its own conclusion before reporting]

The result is an analysis where the confirmatory claims are protected from p-hacking and HARKing, the anomalies were explained rather than hidden, and every number can be reproduced.

What's inside

Skills library

Framing

• framing-research-questions — Turn an interest into a falsifiable question (entry gate)
• surveying-prior-work — Ground the question and methods in existing literature
• establishing-feasibility-first — Opt-in exploratory/feasibility mode: prove the computation runs at scale before paying for process

Planning & pre-registration

• designing-the-analysis — Detailed, bite-sized analysis plan
• preregistering-analysis — Lock predictions and decision rules before seeing outcomes (includes a statistical-fallacies reference and prereg.sh, a zero-dependency freeze/audit script that proves from git history alone that predictions preceded outputs and were never edited after the freeze)

Execution

• subagent-driven-analysis — Fresh subagent per analysis step with two-stage review
• executing-analysis — Inline batch execution with checkpoints
• dispatching-parallel-investigations — Concurrent independent investigations

Discipline

• investigating-anomalous-results — 4-phase root-cause process for surprising results
• verifying-results-before-claiming — Evidence before claims

Review

• requesting-red-team-review — Dispatch a skeptical reviewer to attack the analysis
• receiving-critical-review — Respond to critique with rigor, not performative agreement

Workspace & reporting

• setting-up-reproducible-analysis — Isolated, reproducible workspace
• reporting-and-archiving-findings — Decide how to report; archive code, data, environment

Meta

• writing-science-skills — Create new skills following the testing methodology
• using-science-superpowers — Introduction to the skills system

Philosophy

• Pre-registration — State predictions and decision rules before seeing outcomes
• Confirmatory vs. exploratory — Always labeled, never blurred
• Reproducibility — Pinned environments, fixed seeds, immutable raw data
• Evidence over claims — Verify before declaring a finding
• Root cause over patching — Investigate anomalies; don't quietly drop data

From the blog

Essays from the K-Dense blog on this methodology and the problems it is built to solve.

This project

• Introducing Science Superpowers (May 28, 2026) — Why we built it, the Iron Law, and the full workflow: pre-registration over TDD.
• Your AI Assistant Reasons Like a Generalist. Science Needs a Specialist. (June 2, 2026) — Companion open-source AGENTS.md profiles; names Science Superpowers as the methodology layer that puts pre-registration ahead of trial and error.

Why discipline, not a smarter model

• AI Co-Scientist, Not AI Scientist: Why the Name Matters (May 5, 2026) — Keep the human scientist front and center; the agent is a partner, not a replacement.
• The AI Co-Scientist Is Here. The Bottleneck Is Verification. (June 3, 2026) — Power is not the open question; whether the work is verifiable is.
• The Model Is No Longer the Bottleneck (June 7, 2026) — Capability has moved; the remaining bottleneck is the workflow around the model.
• The Week Science Models Became Real (June 12, 2026) — Frontier models entered scientific workflows; the next bottleneck is evidence.
• Reproduction, Not Generation, Is AI's Killer App for Science (June 16, 2026) — Agents that reproduce published findings matter more than agents that generate unchecked claims.
• AI Scientists Need Lab Escape Rooms, Not More Exams (June 29, 2026) — Benchmarks should demand real evidence, not science theater.

How agent skills work

• Agent Skills: The Final Piece for AI-Powered Scientific Research (January 13, 2026) — Why composable skills close the gap between raw model intelligence and domain expertise.

Installation

Installation differs by harness. If you use more than one, install Science Superpowers separately for each.

Agent Plugins (any conformant client)

This repository is a valid Agent Plugins v1.0.0 package — the open, vendor-neutral plugin standard. The portable manifest is plugin.json at the repository root, and the sixteen skills are the immediate children of skills/, each with its own SKILL.md.

Any client that implements the specification can install Science Superpowers by pointing at this repository; no harness-specific configuration is required for the skills themselves. Bootstrap hooks remain client-specific — see the sections below for the harness you use, or invoke the using-science-superpowers skill manually at the start of a session.

Cursor

In Cursor Agent chat, install from the plugin marketplace, or point Cursor at this repository as a plugin. The sessionStart hook (hooks/hooks-cursor.json) loads the bootstrap automatically.

Claude Code

Register a marketplace pointing at this repo (.claude-plugin/marketplace.json) and install the science-superpowers plugin. The SessionStart hook (hooks/hooks.json) loads the bootstrap.

Codex

Use the committed Codex manifest at .codex-plugin/plugin.json.

Gemini CLI

Install as an extension; gemini-extension.json points the context file at GEMINI.md, which loads the bootstrap and the Gemini tool mapping.

OpenCode

See .opencode/INSTALL.md.

Pi

Pi natively reads the same SKILL.md format. Install the package with pi install git:github.com/K-Dense-AI/science-superpowers; the package.json pi key registers the skills and a before_agent_start extension loads the bootstrap. See .pi/INSTALL.md.

Google Antigravity

Antigravity natively supports Agent Skills (the same SKILL.md format) and reads GEMINI.md / AGENTS.md / .agent/rules/ as always-on rules at session start. Install the skills and load the bootstrap rule — see .antigravity/INSTALL.md.

Contributing

See AGENTS.md / CLAUDE.md for contributor guidelines, and skills/writing-science-skills/SKILL.md for the complete guide to creating and testing skills.

License

MIT License — see the LICENSE file. This project reimplements the architecture of Superpowers by Jesse Vincent.

Star History