Scientific Research

Codeg bundles a curated library of scientific-research skills that turn any of its coding agents into a research assistant. They carry a study across its whole arc — from a half-formed idea to a testable hypothesis, a rigorous experimental design, an honest power calculation, real statistical analysis, publication-ready figures, and a search of the literature. You invoke them the same way you invoke any skill: type / and pick one.

Where the Office pack produces research artifacts — the deck, the paper, the workbook — the Science pack supplies the method: how to frame a question, design the study, run the stats, and appraise the result. The two compose well; this page is about the thinking half.

The skills are vendored, byte-for-byte, from the open-source K-Dense-AI/scientific-agent-skills project (MIT-licensed), baked into Codeg, and enabled per agent like any pack.

Enable the research skills

The Science pack lives under Settings → Skill Packs → Science. It shares the same central store and enabling model as the other packs: a skill-and-agent matrix where you tick which skills each of your agents gets, and Codeg links them into that agent's own skills directory. Nothing here is agent-specific magic — a linked science skill is just a skill in that agent's / menu.

Two badges show up on the Science matrix that the Experts pack doesn't have — they flag a skill whose own workflow needs something before it runs:

• May need setup — the skill ships helper scripts that expect a Python (often uv) environment. It still links fine; the script-backed steps just need that environment present. The skill's own SKILL.md lists what to install.
• Needs API key — the skill's main workflow calls an external service that wants a key. Only one skill carries this (Scientific Schematics — see below); it links like any other, but the diagram step needs the key set first.

Enabling always works — setup is about the skill, not the link

A May need setup or Needs API key badge never blocks enabling. It's a heads-up that the skill's tools have a prerequisite. Keys go where the agent's other credentials live — an environment variable the agent runs with. → Where credentials are stored

The research skills

Thirteen skills, grouped by the stage of research they serve. Each has its own /-invocation and knows the conventions of its craft:

Stage	Skill	What it does	Setup
Ideation	Scientific Brainstorming	Explore interdisciplinary links, challenge assumptions, surface research gaps	—
	Hypothesis Generation	Turn observations into testable hypotheses with predictions and mechanisms	Python
Study Design	Experimental Design	Randomization, blocking, controls, and factorial/DOE layouts — before you collect data	Python
	Statistical Power	Sample-size and power analysis, minimum detectable effects, power curves	Python
Analysis	Statistical Analysis	Test selection, assumption checks, effect sizes, APA-style reporting	Python
	Exploratory Data Analysis	Automated exploration of data files across 200+ formats, with quality reports	Python
Visualization	Scientific Visualization	Publication-ready figures — multi-panel layouts, significance annotations, journal formats	Python
	Scientific Schematics	AI-generated diagrams — pathways, architectures, flowcharts	API key
Evaluation	Critical Thinking	Appraise claims and evidence — biases, confounders, GRADE and risk-of-bias	—
	Peer Review	Checklist-based manuscript and grant review, CONSORT/STROBE compliance	Python
	Scholar Evaluation	Score work with the ScholarEval framework — formulation through writing	Python
Literature	Paper Lookup	Search 10 scholarly APIs (PubMed, arXiv, OpenAlex, Crossref…) for papers and full text	—
	Citation Management	Verify metadata and generate accurate BibTeX — DOI-to-BibTeX, reference validation	Python

Two are worth calling out. Paper Lookup reaches ten open scholarly APIs and needs no key — it's a keyless literature search you can enable and use immediately. Scientific Schematics is the one skill that needs an API key: it renders diagrams through an AI image model and expects an OpenRouter key.

You only enable the ones a given agent will use — a data-analysis agent wants the stats and EDA skills, not the pitch for peer review.

Start from the welcome screen

A fresh conversation shows quick-action tabs; the Scientific Research tab is the fast way in. It promotes the first three skills — Scientific Brainstorming, Hypothesis Generation, Experimental Design — as prominent cards, with the rest of the featured skills in a scrolling row. Clicking one drops the skill in and seeds a starter prompt you finish in your own words:

/scientific-brainstorming Help me brainstorm research directions — explore
interdisciplinary connections, challenge assumptions, and surface promising
gaps. The area I'm exploring: circadian regulation of gut microbiota

If a skill isn't enabled for the current agent, its card shows a lock — click it and Codeg points you to Settings to switch it on. Prefer to type? / in the composer lists every enabled skill, so /paper-lookup or /statistical-analysis works without touching the cards. (Codex uses $ instead of /.)

The four skills that aren't featured — Peer Review, Citation Management, Scholar Evaluation, Scientific Schematics — don't get welcome cards, but they're right there in the / menu and the settings matrix once enabled.

A research session, end to end

The skills are built to hand off to one another, so a single conversation can walk the whole lifecycle:

1. Frame it. /scientific-brainstorming to widen the question, then /hypothesis-generation to sharpen it into something with a prediction and a mechanism.
2. Design it. /experimental-design lays out controls and randomization; /statistical-power tells you how many samples you actually need before you spend a week collecting them.
3. Analyze it. Point /exploratory-data-analysis at your data file for a first pass, then /statistical-analysis to pick the right test, check its assumptions, and report effect sizes.
4. Show it. /scientific-visualization renders publication-ready figures; /scientific-schematics draws the mechanism diagram.
5. Pressure-test it. /scientific-critical-thinking and /peer-review appraise the claim and the write-up the way a reviewer would; /paper-lookup and /citation-management ground it in the literature and produce clean BibTeX.

The script-backed skills (the Python ones) run their analysis and drop the outputs — reports, figures — straight into your working folder, where they open in a workspace tab. When the science is settled and you want the write-up as a real .docx or a slide deck, hand the results to the Office pack.

Good to know

• "May need setup" is about the skill's scripts, not Codeg. The stats, EDA, visualization, and most evaluation/literature skills ship Python helpers; enabling links them instantly, but those steps run only once a Python/uv environment is available on the machine the agent works on. Each skill's SKILL.md spells out its dependencies.
• One key, one skill. Only Scientific Schematics needs an API key (OpenRouter, for AI-rendered diagrams). Everything else — including the ten-API Paper Lookup — runs without one.
• They stay current on their own. Because the pack is bundled into Codeg and re-extracted to the central store on each upgrade, enabled skills track the vendored upstream with no action from you. Edit one in place and Codeg marks it User modified and preserves your version.
• Enable per agent, invoke like any skill. Same matrix, same /id (Codex $id), same Reconnect to apply for a session that's already open. → Skills · Configure an agent
• It works the same on a server. The Skill Packs screen is in the browser build; the skills live on the machine Codeg runs on.

Next steps

• Skills — the enabling matrix in full, and how to write your own skills.
• Office Documents — turn research results into real .docx, .xlsx, and .pptx files with live preview.
• MCP Servers — give an agent live data tools and APIs beyond the bundled skills.
• Working with Agents — pick and sign in the agent that will do the research.