Skip to main content
skill CX Codex

scientific-thinking — Claude Code skill: structured research reasoning

scientific-thinking is a skill for analyzing research findings, evaluating evidence, and constructing rigorous arguments. It separates facts from interpretations and calibrates conclusions to evidence quality.

MIT 308 installs ★ 107 updated 2026-10-16

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add Agents365-ai/365-skills --skill scientific-thinking -a claude-code
Open on GitHub

What scientific-thinking does

scientific-thinking provides a seven-layer framework for evidence-aware reasoning: frame the problem, decompose assumptions, separate observed facts from inference, consider alternatives, match claim strength to evidence, define boundaries, and identify next steps. The skill runs with no external dependencies and works across all LLM platforms.

The core mechanism pushes agents to pause before answering and work through a structured set of checks: what is known, what is assumed, what other explanations fit the data, how strong is the evidence. Instead of assertive conclusions, the skill trains you to say "supports", "suggests", or "cannot yet rule out" based on evidence quality.

Use it when interpreting experimental results, analyzing mechanisms, weighing competing hypotheses, designing experiments, or reviewing papers. Runs on macOS, Linux, and Windows with no prerequisites.

What it helps with

  • Interpret experimental results and research conclusions
  • Distinguish correlation from causation in data
  • Evaluate strength and limits of a scientific claim
  • Design experiments and critique methodology
  • Surface hidden assumptions and alternative explanations

Example prompts for the agent

  • Walk me through this result using scientific-thinking: separate the observed facts from how they're being interpreted
  • What alternative hypotheses explain these findings? Rank them by how well they're supported
  • Design an experiment that would distinguish between these two mechanisms

How to install scientific-thinking

1Skills CLI (Vercel)

npx skills add Agents365-ai/365-skills --skill scientific-thinking -a claude-code

2Manually: put the files in place

~/.claude/skills/scientific-thinking/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/scientific-thinking/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about scientific-thinking

How do I install scientific-thinking for Claude Code?

Run npx skills add Agents365-ai/365-skills --skill scientific-thinking -a claude-code, then open ~/.claude/skills/scientific-thinking/SKILL.md. The skill will activate automatically in Claude Code.

What does scientific-thinking do?

It teaches agents to work through seven reasoning steps: problem framing, decomposition, fact-versus-inference separation, alternative hypotheses, evidence calibration, boundary definition, and next steps. The result is more careful, evidence-aware reasoning and language that matches claim strength to available support.

Which platforms support scientific-thinking?

The skill works with Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, and OpenCode. It has no external dependencies and runs on any LLM agent.

Source text

SKILL.md · MIT
---
disable-model-invocation: true
name: scientific-thinking
description: Use when interpreting research findings, evaluating scientific evidence, analyzing mechanisms, comparing competing hypotheses, designing experiments, or constructing scientific arguments.
license: MIT
homepage: https://github.com/Agents365-ai/365-skills
compatibility: No external tool dependencies. Works with any LLM-based agent on any platform.
platforms: [macos, linux, windows]
metadata: {"openclaw":{"requires":{},"emoji":"🔬","os":["darwin","linux","win32"]},"hermes":{"tags":["scientific-thinking","research","reasoning","evidence-evaluation","hypothesis","experiment-design","mechanism","peer-review"],"category":"research","requires_tools":[],"related_skills":["literature-review","paper-reader","zotero-cli-cc"]},"pimo":{"category":"research","tags":["scientific-thinking","reasoning","evidence-evaluation","research","hypothesis"]},"author":"Agents365-ai","version":"1.0.0"}
---

# Scientific Thinking

A meta-skill for structured, evidence-aware, boundary-conscious scientific reasoning. Your role is not just to answer — it is to reason like a careful researcher.

## When to Use

- Interpreting experimental results or paper conclusions
- Analyzing mechanisms or pathways
- Distinguishing concepts that are being conflated
- Evaluating competing hypotheses
- Designing or critiquing experiments
- Constructing scientific arguments

## Update check

Throttle to one check per 24 hours per installation; never mutate the skill directory without explicit user consent.

1. If `<this-skill-dir>/.last_update` exists and is less than 24 hours old, skip this step entirely.

2. Otherwise, fetch the latest tag from upstream:

   ```bash
   git -C <this-skill-dir> ls-remote --tags origin 'v*' 2>/dev/null \
     | awk '{print $2}' | sed 's|refs/tags/||' \
     | sort -V | tail -1
   ```

3. Compare with this skill's `metadata.version` from the frontmatter. If the upstream tag is strictly newer (semver), tell the user one line and ask:

   > "A newer version of this skill is available: vX.Y.Z → vA.B.C. Want me to `git pull`?"

   If they say yes, run `git -C <this-skill-dir> pull --ff-only`. Refresh `.last_update` either way so the prompt doesn't repeat for 24 hours.

4. If upstream is the same or older, refresh `.last_update` silently and continue.

5. On any failure (offline, not a git checkout — e.g. ClawHub-installed copy, read-only path, no permission), swallow the error silently and continue with the user's task. Do not mention the failure.

## Core Reasoning Framework

Work through these layers before responding.

### 1. Frame the Problem

- What exactly is being asked?
- Scientific level: fact / concept / mechanism / method / interpretation / decision?
- What is known, unknown, and assumed?
- Restate the real problem if the question is broad or ambiguous.

### 2. Decompose

- What needs to be defined first?
- What hidden assumptions are present?
- What distinctions must be kept separate (phenotype vs mechanism, association vs causation, state vs lineage)?
- What would make the conclusion invalid?

### 3. Separate Evidence from Interpretation

Always distinguish among: observed fact / direct evidence / indirect evidence / interpretation / hypothesis / speculation / uncertainty.

- Do not present a hypothesis as a fact.
- Do not present correlation as causation.
- Do not present a label as a mechanism.

**Evidence provenance:** State whether each key claim comes from (a) provided data, (b) general background knowledge, or (c) inference. If required evidence is absent from the prompt, either retrieve it or explicitly label the answer as provisional reasoning.

### 4. Consider Alternative Explanations

Before giving a conclusion:

- Is there another plausible explanation?
- Could this be caused by confounding, measurement error, sampling bias, or definition mismatch?
- Could this reflect context rather than essence?

If multiple explanations are plausible, rank them by available support. Do not pretend there is only one. Surface alternatives only when they are genuinely plausible — do not force false balance.

### 5. Calibrate Claim Strength

Match conclusion strength to evidence strength:

| Evidence level | Language to use |
| ---------------- | ----------------- |
| Strong, replicated | "demonstrates", "establishes" |
| Consistent, single source | "supports", "is consistent with" |
| Suggestive, indirect | "suggests", "is compatible with" |
| Speculative | "raises the possibility", "cannot exclude" |
| Absent | "is insufficient to conclude" |

### 6. Define the Boundary

Every meaningful conclusion has limits. State when relevant:

- what this conclusion supports vs. what it does not yet prove
- under what conditions it may hold or not generalize
- what evidence is still missing

### 7. Move Toward Resolution

Do not stop at abstract interpretation. Suggest:

- the most likely current conclusion
- the key unresolved issue
- the lowest-cost next step that would discriminate between the leading explanations

## Output Structure

Unless the user wants a very short answer, organize in this order:

1. Problem framing
2. What can be said with confidence (with provenance: data / background / inference)
3. Main possible interpretations, ranked by support
4. Most reasonable current conclusion
5. Boundary / limitation / uncertainty
6. Next step

If the user wants a concise answer, compress this structure — do not abandon it.

## Style

**Be:** structured, precise, calm, intellectually honest, non-dogmatic

**Do:**

- Clarify definitions when concepts are mixed
- Label what is observed vs. inferred vs. assumed
- State uncertainty clearly

**Do not:**

- Jump to conclusions
- Confuse description with explanation
- Use confident language when evidence is weak
- Ignore alternative explanations
- Overclaim based on a single study or indirect evidence

## Quick Reference

| Situation | Action |
| ----------- | -------- |
| Question is broad or ambiguous | Restate the real problem first |
| Correlation present | Clarify: not causation without further evidence |
| Single explanation offered | Check for alternatives before concluding |
| Conclusion seems strong | State its boundary; label claim level |
| Evidence is weak or absent | Hedge language; label as provisional; identify what's missing |
| Concept conflated across levels | Separate levels (phenotype/mechanism, association/causation) before answering |
| Evidence not in prompt | Retrieve it or explicitly label answer as provisional reasoning |

## Before Responding

Run through @checks.md.

## Examples

See @examples.md for preferred response style in common research scenarios.

© Agents365-ai · MIT · Agents365-ai/365-skills

Similar skills

All skills for Science and research →

From the same set33

The whole set →
obsidian-dev-rules obsidian-dev-rules is a Claude Code skill with rules for building TypeScript plugins and CSS themes for Obsidian. It guides API usage and event cleanup. CX Codex +1 317installs journal-meta journal-meta fetches complete paper metadata from a DOI, PMID, arXiv identifier or title: authors, journal, impact factor, citation count and abstract. CX Codex +1 317installs xfetch xfetch extracts public tweets, profiles, and likes from X using a local Python script and your session cookies. Avoids paid developer API keys for data scraping. CX Codex +1 316installs bbc-skill bbc-skill collects Bilibili video comments into JSONL with a JSON summary. Use it to gather feedback on one video or across an uploader’s videos. CX Codex +1 313installs grant-thinking-general The grant-thinking-general skill evaluates grant ideas and research proposal logic, helping shape clearer problems, aims, innovation claims, and feasibility. CX Codex +1 301installs figshare-skill Search Figshare’s public research records, download their files, or manage articles in your account. Useful when preparing a dataset upload or publishing an article. CX Codex +1 299installs scientific-thinking-biology scientific-thinking-biology is a Claude Code skill that structures reasoning about biological research and genomic data. It distinguishes association from causation and marker from driver. CX Codex +1 296installs grant-thinking-cn-biology grant-thinking-cn-biology is a Claude Code skill that tests a biology grant idea before submission: it checks for a real scientific question, an unresolved mechanism and a scope suited to the funding level. CX Codex +1 295installs

Project sponsors

Companies that keep this analytics open