How to Search — The Skill of Finding What You Don't Know

The gap between a strong developer and a beginner is closer to "finds what they don't know quickly" than "knows a lot." Search is not a supplement to memorization; it's a separate skill. This article covers how to craft good keywords, the priority between official docs and community sources, English-language search, and the criteria for using LLMs as search aids.

1. Where search sits

Search engines went through Archie, Yahoo Directory, and AltaVista in the early 1990s, then became the de-facto standard tool after Google PageRank arrived in 1998. For developers, one of Google, DuckDuckGo, Kagi, Brave Search, or Bing is part of daily work. Since 2022, LLMs like ChatGPT, Claude, and Gemini have started absorbing part of the search role.

Search is two stages:

1. Make what you don't know concrete.
2. Convert that ignorance into keywords the search engine can understand.

Most stuck moments come from step 1 being fuzzy.

2. Crafting good keywords

A common mistake is dropping a natural-language sentence straight into the search box. Search engines find the intersection of keywords, so rare nouns and exact error messages work better than sentences.

Weak search	Better search
"node isn't working"	Error: Cannot find module 'express'
"how do I send react props"	react pass props to child component
"python date comparison fails"	python compare datetime naive vs aware TypeError

Paste error messages verbatim where possible. If the message contains variable parts like usernames or file paths, drop them or replace with asterisks.

3. Quotes and operators

Syntax	Effect
"exact phrase"	Match word order inside quotes
-word	Exclude results containing that word
site:stackoverflow.com	Limit to a specific site
site:github.com inurl:issues	Specific site + URL pattern
filetype:pdf	Filter by extension
intitle:node	Match in title

For example:

"TypeError: Cannot read properties of undefined" site:stackoverflow.com
react useState site:react.dev
postgres "deadlock detected" -mysql

4. Priority — where to look first

This order is often recommended:

1. Official docs — react.dev, nodejs.org, postgresql.org and other primary sources. Most accurate, most current.
2. GitHub issues / discussions — maintainer answers from the library itself may be there.
3. Stack Overflow — answers carry scores and acceptance marks. Read the highly-rated answer along with its comments.
4. MDN — the de-facto reference for web standards.
5. Personal blogs — check the author and date. A 5-year-old post may be stale.
6. Reddit, Hacker News, GeekNews — for tone-checking. Don't copy answers verbatim.

To avoid grabbing old answers, use the search engine's "Tools > Past year" option or the &tbs=qdr:y parameter to limit results to the past year.

5. Searching in English

Korean material is small in volume and slow to update. The same question in English yields 5 to 50 times more material. We don't need English writing ability — only the ability to compose English keywords.

Korean keyword	English keyword
"리액트 부모 자식 통신"	react parent child communication
"스프링 부트 cors 설정"	spring boot cors configuration
"포스트그레 인덱스 안 타요"	postgres query not using index

Error messages are often English from the start, so pasting them directly is usually enough.

6. Using LLMs to supplement search

LLMs like ChatGPT, Claude, and Gemini are strong in these spots:

• Turning a vague problem into keywords ("the symptom is X — what would I search for?")
• Summarizing docs and code
• Mapping out trade-offs between several paths
• Generating "5 next things to try" candidates when stuck

The weak spots are clear too:

• They may not know information after the model's training cutoff.
• API and library versions often drift (non-existent functions, wrong signatures).
• They answer with confidence. Pasting into code without verification breaks things.

Treat an LLM's answer as "the opinion of one confident friend," and verify code or important decisions against the official docs. Modes that combine web search (GPT-4 Browse, Claude with web, Perplexity) make verification easier by providing source links.

7. Other paths

• Perplexity, Phind, You.com — search that returns LLM answers with source links.
• DevDocs.io — multiple official docs in one offline interface.
• Kagi — a paid search with less ad and SEO spam.
• GitHub Code Search — pattern search directly across open source code.
• DuckDuckGo !bang — !so query for site-specific search shortcuts.

8. Stuck → search → answer flow

1) Red error on screen: "TypeError: Cannot read properties of null (reading 'map')"
2) Refine keywords: drop variable parts (filename, line number)
3) Search: "Cannot read properties of null (reading 'map')" react
4) First result — Stack Overflow: .map called before fetch finishes → null guard
5) Verify against official docs — react.dev "Conditional rendering"
6) Fix code: items?.map(...) or items && items.map(...)

9. Common pitfalls

Searching without reading the error message — more than half the time the message itself is the answer. Read it calmly.

Keywords too long — five or more words in one search returns almost nothing. Keep only the core.

Adopting old answers — applying a 2017 React answer in 2026. Always check the date and library version.

The top-voted Stack Overflow answer isn't always right — a better answer often hides in the comments.

Search refusal — "this is too basic to search for" is a big trap. Every senior searches every day.

Trusting LLM answers verbatim — hallucinations are routine. Check whether sources exist and whether it actually works.

Closing thoughts

Search is not a supplement to memorization but a separate skill. The five places — good keywords (rare nouns, exact error messages), priority (official docs → GitHub → Stack Overflow), English keywords, LLM assistance, and verification — together unstick problems fastest. We live in an age where every senior searches every day.

Next

• how-to-read-docs
• how-to-ask-good-questions

Google Search Operators official guide · Google Programmable Search · DevDocs · Kagi · GitHub Code Search · Perplexity · Phind · Claude · ChatGPT · Gemini · Stack Overflow · GeekNews (Korea) for reference.