Stop wrestling the search box.
Tell Search Refiner what you’re looking for in plain words. It writes the search query for you the way a librarian would, using the exact phrases and operators that actually surface good sources. You get there in one or two tries instead of ten.
New to skills? A Claude skill is a short set of written instructions you add to Claude once, so it handles one kind of task well. This one writes search queries for you.
Good searching is a craft — most of us were never taught it.
The open web rewards precise queries and buries lazy ones under SEO’d affiliate spam. Search Refiner does what a good reference librarian does out of habit: it cuts your question down to its core ideas, finds the words experts actually use, and chains the right operators together.
Searching by hand
- You type a whole sentence and Google matches the words separately.
- Page one is affiliate “best of” lists, not the primary source.
- You don’t know the term the experts use, so the good results never surface.
- You retry five, six, ten times, tweaking words at random.
- You trust the first plausible-looking site without checking who’s behind it.
Searching with the skill
- Exact phrases get quoted; synonyms get OR-grouped so nothing is missed.
site:,filetype:and date bounds cut straight to PDFs, .gov, .edu, or a fresh window.- It suggests the specialist vocabulary you didn’t know to use.
- A tight query first, then a looser fallback — the precision ladder.
- A built-in trust check: how to verify an unfamiliar source before you cite it.
Any time you’re searching the open web or scholarly literature.
It’s for finding things on the open web, not for questions a chatbot can answer on its own.
Reach for it when…
You don’t need it when…
Three moves, borrowed from the reference desk.
You describe what you’re looking for, in plain words. The skill then runs the same routine a search librarian does.
Understand what you’re after
It strips your question to 2–4 core concepts expressed as nouns, then brainstorms the synonyms and specialist terms experts use. It also reads what form you need (fact, study, PDF, dataset, image origin), whether freshness matters, and how high your trust bar is.
“teens” → adolescent*, youth, “high school students”Build the queries
It composes a query from the verified operator set — quoted exact phrases, site:, -exclude, filetype:, before:/after:, capitalized OR with grouping — and never suggests dead operators like ~, +, or cache:. The power is in chaining them.
Deliver, ready to paste
You get a best query as a copy-paste string plus a clickable link, 2–3 variants (a broader fallback, a scoped version, sometimes a different engine), a one-line note on where to run it, and a trust check — the fastest way to vet an unfamiliar source before you rely on it.
Tight version first → looser fallback → verify laterallyThe Operator Lab
Feel how a query is built. Type a concept or two, flip a few operators, and watch a ready-to-paste string assemble. This is the move the skill makes for you — here you get to make it by hand.
Compose a search
Everything you toggle appears in the live query below. Nothing is sent anywhere until you click through to the search engine yourself.
The librarian’s move: name the term an expert would use. teens → adolescent*, youth · AI in schools → "generative AI", K-12. Finding the right word is half the search.
Reading the query: quotes force an exact phrase, OR (always capitalized) accepts either term, and every operator hugs its colon with no space. On Scholar, use the “Cited by” link under a strong result to follow the research trail. This lab covers the everyday operators; the skill also handles vocabulary discovery, the precision ladder, and source verification.
Finding it is half the job. Trusting it is the other half.
A sharp query still lands you on sources you’ve never heard of. The first move is the simplest — stop, and notice your own reaction before you trust or share. Then run the three moves fact-checkers use; together they take under a minute. The skill folds these into every result, but they’re worth knowing by hand.
Read laterally
Don’t judge a site by its own About page. Open a new tab and search the source’s name — see what everyone else says about it. Fastest version: search [site name] wikipedia.
Find better coverage
If the source looks shaky, don’t study it harder. Search the claim itself and see whether trusted outlets — or Snopes, PolitiFact — cover it. Better coverage beats a closer read of a weak page.
Trace to the original
Follow quotes, stats, and images upstream to where they first appeared — and check they’re represented fairly, not cherry-picked. For a photo, reverse-search it with Google Lens.
These are the SIFT moves — Stop, Investigate the source, Find better coverage, Trace to the original — the same habits taught in information-literacy classrooms.
What is a skill, exactly?
A skill is a small folder Claude can pick up on its own when a task calls for it. No plugin to install, no button to press — just plain-text instructions plus any reference files. Here’s the actual Search Refiner skill, dissected.
The top block & name
MetadataThe fenced block at the very top (the --- lines) holds the skill’s settings. The name is its handle — lowercase-with-hyphens, like search-refiner — and it names the folder the skill lives in. Everything Claude needs to decide whether to use the skill sits up here.
Description — the trigger
Most importantThis is the load-bearing part. It tells Claude exactly when to reach for the skill: the phrases, situations, and complaints that should wake it up (“refine my search,” “help me find,” “my searches return junk”). A vague description never fires; a precise one fires at the right moment.
The body — the method
InstructionsPlain Markdown telling Claude how to do the work: the persona (“expert search librarian”), the numbered steps, the rules that matter most, the required output format, and a worked example. This is the craft, written down.
Reference files
Loaded on demandBulky details live in separate files — here, operators.md holds the full verified operator tables and engine notes. Claude opens them only when needed, so the main instructions stay lean. This is called progressive disclosure.
Why the shape matters
The ideaA good skill is model-invoked: you don’t summon it by name, Claude notices the moment fits and picks it up. That only works if the description is honest about when it applies — which is why “when to use it” is the part worth sweating.
It’s just text. You can read every word before you paste it.
Nothing to download. Nothing runs on your computer. A skill is plain-language instructions Claude reads — no app, no code, no installer. Everything it does is right here in the open. Copy it, read it if you like, and paste it in.
---
name: search-refiner
description: Turn a plain-language description of what someone is hunting for into expertly crafted, ready-to-paste search queries for Google, Google Scholar, and other search engines — using the operator syntax and query-construction methods that librarians and information literacy specialists teach. Use this skill whenever the user wants to search for something online, asks "how should I google this," wants help finding research, articles, PDFs, data, or primary sources, says "refine my search," "build me a search query," "help me find," or describes something they've been struggling to locate. Also trigger when the user complains that their searches return junk, wants to search within a specific site or date range, or asks which search engine to use. If the user is trying to FIND information on the open web or in scholarly literature, use this skill.
---
# Search Refiner
You are acting as an expert search librarian. The user describes what they're hunting for in plain language; you hand back optimized, ready-to-paste search queries plus a quick note on where to run them. The goal is that they paste one string and land on the right results in one or two tries instead of ten.
## Step 1 — Understand the hunt (fast)
Extract from the user's request (ask at most 2-3 quick questions ONLY if you truly can't infer these):
1. **Core concepts** — the 2-4 main ideas, expressed as nouns. This is the librarian's first move: strip the question down to concept nouns, then brainstorm synonyms and the specialized terms experts in that field would use. ("teens" → adolescent*, youth, "high school students"; "AI in schools" → "generative AI", "artificial intelligence", K-12, classroom)
2. **Desired form** — a fact, a news story, a scholarly study, a PDF report, a dataset, a person/org background check, an image origin?
3. **Freshness** — does recency matter? (drives before:/after: and date filters)
4. **Trust bar** — casual answer vs. citable source? (drives site: scoping and engine choice)
## Step 2 — Build the queries
Construct queries with the current, verified operator set. Full tables, syntax rules, and engine notes live in `references/operators.md` — read it before building queries.
The rules that matter most:
- **Quote exact phrases**: `"generative AI" "middle school"` — multi-word concepts must be quoted or the words are matched separately.
- **No space after the colon**: `site:nytimes.com` works; `site: nytimes.com` does not (Google documents this explicitly).
- **The reliable core set**: `"exact phrase"`, `site:`, `-exclude`, `filetype:`, `before:YYYY-MM-DD`, `after:YYYY-MM-DD` — these six are officially documented by Google. `intitle:`, `inurl:`, `intext:`, `OR`, `*` wildcard, and `( )` grouping still work but are third-party-verified, not officially documented.
- **Google implies AND** between every term — never type AND in Google. Type `OR` (capitalized) to accept synonyms, and group them: `("screen time" OR "device use") teens sleep site:.edu`
- **Chain operators** — the power comes from combining: `"professional development" filetype:pdf site:.org after:2024-01-01`
- **Never use deprecated operators**: `~`, `+`, `link:`, `info:`, `cache:`, `daterange:`, `allinurl:` are dead or broken. `related:` and `AROUND(x)` are unreliable. Details in the reference file.
- **Precision ladder**: offer a tight version first, then a looser fallback. Quoted phrases and site: restrictions can over-narrow; the trade is precision vs. recall, and searching is iterative by design.
## Step 3 — Deliver in this format
ALWAYS use this exact structure:
**Best query** — one ready-to-paste string in a code block, followed by the full clickable URL form (`https://www.google.com/search?q=` + URL-encoded query) so the user can click straight through.
**Variants** — 2-3 alternates: a broader fallback (fewer quotes/restrictions, OR'd synonyms), a scoped version (site:/filetype:/date-bounded), and where useful a different-engine version.
**Where to run it** — one line. Default Google; Google Scholar for scholarly work (note: Boolean operators must be CAPITALIZED there; use "Cited by" to follow research trails); Google Images/Lens for image verification; a privacy or non-AI engine (DuckDuckGo, Startpage, Kagi) when the user wants results without AI overviews. Engine details in the reference file.
**Trust check** — one or two lines applying SIFT/lateral reading: if the results will come from unfamiliar sources, remind the user of the fastest verification move for their case. The best single trick: search `[domain name] wikipedia` to get instant third-party context on an unfamiliar site. For contested claims, suggest finding better coverage (search the CLAIM, not the site) rather than reading the suspect page harder. For images, reverse image search via Google Lens with the Tools date filter.
Keep the whole response tight — queries first, commentary minimal. The user wants strings to paste, not an essay.
## Example
**Input**: "I want recent research on how phone bans affect learning, not news fluff"
**Output**:
Best query (Google Scholar):
```
"phone ban" OR "smartphone ban" schools "academic achievement"
```
https://scholar.google.com/scholar?q=%22phone+ban%22+OR+%22smartphone+ban%22+schools+%22academic+achievement%22
Then filter with the sidebar "Since 2024" and chase the best paper's "Cited by" list.
Variants:
- Broader (Google): `smartphone ban schools effect learning after:2024-01-01 -site:reddit.com`
- Policy PDFs: `"mobile phone ban" schools filetype:pdf site:.edu OR site:.gov`
Trust check: Scholar indexes some non-peer-reviewed and predatory-journal content — check the journal name laterally (`[journal name] predatory`) before citing.
## Iterating
If the user comes back with "that didn't work": diagnose which failure mode — too few results (drop quotes, add OR synonyms, remove site:), too much junk (add quotes, exclude with -, scope with site: or filetype:), or wrong vocabulary (mine the near-miss results for the terms experts actually use, then rebuild — librarians call this tracking discovered vocabulary, and it's the single most effective iteration move).# Operator Reference & Engine Guide (verified July 2026)
Sources were fetched and quote-checked in July 2026. Confidence labels: **[official]** = Google's own documentation; **[verified]** = consistent across multiple current third-party references (Ahrefs, Search Engine Land, 2025 operator audits); **[library]** = librarian-taught technique from university library guides (MIT, UChicago, UConn, SUU, FDU, Illinois).
## Google Search — operators that work
| Operator | What it does | Example | Confidence |
|---|---|---|---|
| `"..."` | Exact phrase match (capped ~32 words) | `"project based learning"` | official |
| `site:` | Restrict to domain/URL prefix; works with TLDs | `site:edutopia.org`, `site:.gov` | official |
| `-` | Exclude a term or site | `jaguar -car`, `-site:pinterest.com` | official |
| `filetype:` | Restrict to file format (pdf, docx, pptx, xlsx, csv) | `rubric filetype:pdf` | official |
| `before:` / `after:` | Date bounds, YYYY-MM-DD; combinable | `after:2025-01-01 before:2025-06-30` | official |
| `imagesize:` | Exact pixel dimensions (Google Images only) | `imagesize:1920x1080` | official |
| `OR` (or `\|`) | Either term; MUST be capitalized | `("AI literacy" OR "digital literacy")` | verified |
| `( )` | Group OR'd synonyms | `(teens OR adolescents) "screen time"` | verified |
| `*` | Wildcard placeholder inside phrases | `"the * of education"` | verified |
| `intitle:` | Term in page title | `intitle:syllabus "media literacy"` | verified |
| `inurl:` | Term in URL | `inurl:handbook` | verified |
| `intext:` | Term in body text | `intext:"logic model"` | verified |
| `source:` | Restrict to a publication (Google News) | `source:the_atlantic` | verified |
**Syntax rules (Google-documented):**
- No space between operator and argument: `site:nytimes.com` ✓, `site: nytimes.com` ✗
- Google inserts an implicit AND between all terms — never type AND
- Operators chain: `"teacher retention" filetype:pdf site:.org after:2024-01-01`
- `allintitle:` / `allintext:` variants must be used ALONE — they break when combined with other operators
## Deprecated / broken — never suggest these
| Operator | Status |
|---|---|
| `+` (exact match) | dropped 2011 |
| `~` (synonyms) | dropped 2013 |
| `link:` | dropped 2017 |
| `info:` / `id:` | dropped 2017 |
| `cache:` | retired along with Google's cache feature |
| `daterange:` | obsolete — use before:/after: or the Tools menu |
| `allinurl:` | partially broken, inconsistent |
| `related:` | unreliable; only semi-works for very large sites |
| `AROUND(x)`, `inanchor:`, `#..#` ranges, `loc:` | inconsistent — use with caution, warn the user |
## Google Scholar
- Boolean `AND` / `OR` / `NOT` work but **must be capitalized** (search itself is case-insensitive)
- `author:"d knuth"` finds papers by author; exact title in quotes locates a specific paper
- `source:` restricts to a publication
- **"Cited by"** under each result = forward citation chasing — the core expert move for following a research trail and gauging influence
- Sidebar: "Since [year]" for recent work (relevance-sorted); "Sort by date" for newest
- Advanced search dialog lives under the hamburger (☰) menu
- Every search caps at 1,000 results — precision beats scrolling
- Free alerts by email (no Google account needed)
- Caveat to pass on: Scholar includes non-peer-reviewed and predatory-journal content; the index (~160M docs, unofficial estimate) has no published coverage list. Vet the venue laterally before citing.
## Verticals & image verification
- **Google Images / Lens**: reverse image search via images.google.com camera icon, right-click → "Search image with Google," or upload/paste URL. Lens shows "Exact Matches," "Visual Matches," and "About this image" (image history + whether fact-checkers have flagged it). Filter by date via Tools to find first publication — the standard journalist technique for spotting recycled/miscaptioned images.
- **Google News**: `source:` targeting; date tools.
- **Google Books**: full-text search of print — good for pre-internet claims and quote-tracing.
## Other engines — when to recommend
| Engine | When |
|---|---|
| **DuckDuckGo** | Privacy; `noai.duckduckgo.com` for a fully AI-free results page |
| **Startpage** | Google-quality results proxied without tracking or AI overviews |
| **Brave Search** | Independent index (rare) — useful second opinion when Google's results feel homogeneous |
| **Mojeek** | Fully independent UK index, one-click AI disable |
| **Kagi** | Paid, ad-free, no tracking, AI only on demand — for heavy researchers |
| **Wikipedia (directly)** | Fastest lateral-reading tool for background on any org, publication, or author |
Honest caveat from 2024-2025 research (Leipzig/Bauhaus-Weimar year-long study of 7,392 queries): for product/commercial queries, Google, Bing, AND DuckDuckGo are all dominated by SEO'd affiliate content — switching engines doesn't fix that. Fix it with query craft instead: `-review`, `-best`, `site:reddit.com` for real user opinions, or date-bound to pre-spam eras.
## The verification layer (SIFT — Mike Caulfield, 2017)
Four moves, taught by librarians everywhere; the skill's Trust check draws on these:
1. **Stop** — before trusting/sharing, ask: do I know this source?
2. **Investigate the source** — read *laterally*: open a new tab and search what others say about the site, don't read the site's own About page. Fastest version: `[domain] wikipedia`.
3. **Find better coverage** — if the source is weak, don't keep evaluating it; search the CLAIM and find trusted coverage (FactCheck.org, Snopes, PolitiFact, WaPo Fact Checker).
4. **Trace to the original** — follow quotes, stats, and images upstream; check they're fairly represented, not cherry-picked.
Evidence this works: six 50-minute lessons of Stanford's lateral-reading curriculum doubled high schoolers' ability to spot dubious sites (Journal of Educational Psychology, 2022). Professional fact-checkers beat historians and PhDs at source evaluation precisely because they read laterally instead of closely.
## Key sources
- Google operator docs: https://developers.google.com/search/docs/monitor-debug/search-operators and https://support.google.com/websearch/answer/2466433
- Ahrefs operator list: https://ahrefs.com/blog/google-advanced-search-operators/
- Search Engine Land (19 working operators, late 2024): https://searchengineland.com/advanced-google-search-operators-388355
- Scholar help: https://scholar.google.com/intl/en/scholar/help.html + https://paperpile.com/g/google-scholar-guide/
- SIFT: https://mikecaulfield.substack.com/p/sift-for-ai-introduction-and-pedagogy + https://guides.lib.uchicago.edu/c.php?g=1241077&p=9082322
- Lateral reading study: https://ed.stanford.edu/news/it-doesn-t-take-long-learn-how-spot-misinformation-online-stanford-study-finds
- Keyword generation: https://guides.lib.uconn.edu/findinformation/keywords + https://library.suu.edu/LibraryResearch/keywords-and-boolean
- Boolean logic: https://libguides.mit.edu/c.php?g=175963&p=1158594
- Reverse image search: https://newsinitiative.withgoogle.com/resources/trainings/reverse-image-search-verifying-photos/ + https://libguides.princeton.edu/medialiteracy/reverseimagesearching
- Engine-quality study: https://www.404media.co/google-search-really-has-gotten-worse-researchers-find/ + https://link.springer.com/chapter/10.1007/978-3-031-56063-7_4Two files make one skill. In Claude, create a new skill using the first block as SKILL.md, then add the second as references/operators.md. New to skills? The zero-setup path: paste both blocks into a fresh chat and ask your search question right after — it works for a one-off.
How to actually use it
- Open Claude and start a normal conversation.
- Describe what you’re looking for, in plain words — and what would count as a good source.
- You don’t name the skill. When your ask is a search, Claude reaches for Search Refiner on its own and hands back ready-to-paste queries.
Once you’ve added the skill above, there’s nothing to switch on — it activates itself when the moment fits.
Put it to work in one sentence.
You never have to name the skill. Just say what you’re looking for — the more you tell it about the kind of source, how recent it should be, and how much you need to trust it, the sharper the queries come back.
- A best query, ready to paste, plus a clickable link
- 2–3 variants — a broader fallback and a scoped version
- One line on where to run it (Google, Scholar, or a no-AI engine)
- A trust check for vetting any unfamiliar source