I made a skill for checking everyday science claims

A paper says 85% of people reported improvement after using blue-light glasses. Sounds like an easy purchase. Then you read the methods: everyone surveyed already used the glasses. There was no clear-lens control group.

That number is in a real 2025 study. It describes what respondents remembered feeling. How much of the improvement came from the filter? The study cannot tell us.

This is the kind of checking I made Science Research for. It is a Markdown skill for Claude, ChatGPT and Codex. It gives the assistant a research process to follow whenever you ask an everyday science question.

One question, with the answer

Use science-research.
What does research say about blue-light glasses for screen-related eye strain?

The live check produced this conclusion:

Blue-light glasses have not shown a reliable extra benefit over clear lenses for short-term screen-related eye strain. The controlled evidence is limited, so this does not settle whether a particular person might find a pair more comfortable.

The Cochrane review contains 17 randomized trials across several outcomes. Only three assessed subjective visual fatigue. That distinction matters when the question is specifically about tired eyes after screen use.

Its largest relevant trial compared blue-blocking and clear placebo lenses during a two-hour computer task. No significant symptom difference was found. The review rated the fatigue evidence low certainty; it leaves plenty of uncertainty about longer use.

The newer favorable survey belongs in the answer too. But its lack of a control group prevents treating reported improvement as a measured filter benefit. Reading further also found inconsistent minimum durations of use between sections of that paper. The notes keep that problem visible instead of silently choosing whichever version fits.

The Science Research page has a compact example and setup instructions.

Compared with a normal ChatGPT or Claude answer

A quick explanation could say:

Evidence for relieving screen-related eye strain is limited. A Cochrane review found little reason to expect a short-term benefit over clear lenses. Some people may prefer the glasses, but longer-term benefits remain uncertain.

That is illustrative wording, rather than a captured response from either product. It gives a reasonable answer. The skill run reaches a similar conclusion, then does more of the work you would need to assess a new claim yourself.

A quick explanation	What this skill run adds
Gives the takeaway	Explains why the favorable survey cannot isolate a filter effect
Links a review	Checks the relevant outcome, search cutoff and overlap with the original trial
Summarizes the research	Records what was actually read, exclusions and unresolved gaps
Mentions personal comfort	Keeps readable forum experiences in a separate report

For a casual question, the short answer may be enough. I would use the skill when I need to inspect conflicting evidence or repeat the same checks on several questions. ChatGPT Deep research and Claude Research can already investigate sources in depth. The reason to add this skill is to keep a particular source policy and appraisal process ready to use.

What the skill added to the work

The useful output was the explanation of why these findings can coexist. Someone can feel better wearing glasses while a controlled trial finds no dependable added benefit from blue filtering. Different questions, different comparisons.

Several checks made that explanation easier to inspect:

• Read beyond the abstract. The survey's methods and limitations explain why its favorable percentage cannot settle the purchase question.
• Check the actual outcome. The review's overall trial count covers more than eye strain. Sleep findings cannot stand in for fatigue results.
• Follow the evidence back. The original trial is already inside the review. Linking both gives detail, without adding an independent experiment.
• Look for newer and contrary work. The survey stayed in the discussion despite disagreeing with the controlled evidence. Unverified newer candidates were recorded as gaps.

There were practical access problems too. Some page fetches returned a browser check or a footer. Browser reading recovered the key full texts; the original trial remained abstract-only. Those are different levels of access, and the notes say which was used.

The skill asks for this work up front. I do not want to remember to request a control-group check, publication-status check and contrary-evidence search every time I ask a question.

Keep the experiences, with their limits

The run also read Reddit and Hacker News accounts. Some wearers described relief; others felt no difference or disliked the tint. One favorable account included other eye-care measures, and one negative account involved changes to other lens features. Neither isolates the filter.

A reply promoting the writer's own eyewear product was excluded. That is a fairly useful distinction to make before presenting a forum thread as a collection of customer experiences.

These accounts went into their own report. They help describe comfort and annoyances, while the scientific confidence rating stays tied to the research.

You get three linked outputs by default: a short scientific brief, detailed research notes and a personal-experience report. Add science only to skip the last one. The source policy is editable Markdown, so you can change the eligible sources for your field.

Install it

Keep the whole science-research folder, including references/. Copying only SKILL.md leaves out the source rules and report instructions.

For a fresh Claude Code installation on macOS or Linux:

git clone --depth 1 https://github.com/eduard-s/skills.git ai-skills
mkdir -p ~/.claude/skills
cp -R ai-skills/skills/science-research ~/.claude/skills/

Then invoke /science-research with your question. These are the personal skill location and invocation described in the Claude Code documentation.

In Claude's web interface, ZIP the individual science-research folder and use the custom skill upload. Claude requires its code execution and file creation setting for Skills, even though this particular package contains only Markdown.

For ChatGPT, follow the official skill guide for the surface available in your account. Standalone skills are supported in the desktop app; web and mobile distribution uses plugins. This repository currently ships a standalone skill folder. You can also attach its Markdown files to a conversation and ask the assistant to follow them for that chat.

The Science Research page has the folder layout, a Codex option and example requests in one place.

What this example establishes

An ordinary chat with a careful prompt can do the same work. The benefit here is having the research requirements saved, reusable and open to inspection. This run shows concrete checks and their output; it does not measure an accuracy advantage over another prompt.

The assistant still needs browsing and has to follow the instructions. A narrow source list can miss useful papers. Native Claude and ChatGPT imports remain untested; the evaluation notes record what has actually been checked.

For me, the useful result is being able to open the answer and immediately see why an impressive-looking number carries less weight than it first appears to.