websearch-multilang

name

websearch-multilang

description

Use when the task is to FIND ideas, solutions, root causes, workarounds, prior art, or benchmarks ONLINE — researching a problem/idea/solution on the web — NOT a simple single-doc API lookup. Searches across MANY languages in parallel: spawns one research subagent per language (English, Chinese Simplified, Chinese Traditional, French, Russian, Hebrew, Spanish), each searching native technical terms and native forums, then synthesizes. Triggers: "research this", "search online for", "find a solution/idea/approach", "how do people solve", "is there prior art", "find workarounds", "deep multilingual web research".

websearch-multilang — Parallel Multilingual Web Research

Use this instead of a plain WebSearch whenever the goal is to discover ideas, solutions, root causes, workarounds, prior art, or benchmarks. The best fix for a bug or the best idea for a design is often documented only in Chinese, Russian, or Hebrew. English-only search misses it.

Do NOT use for a simple "what does this API param do" single-doc lookup — a direct WebSearch/WebFetch is enough there.

Flow

digraph websearch_multilang {
    "Frame the problem + stack" [shape=box];
    "Spawn one subagent per language (parallel)" [shape=box];
    "Each: native terms -> WebSearch -> WebFetch strongest" [shape=box];
    "Each: return structured findings" [shape=box];
    "Coordinator synthesis" [shape=box];
    "Final answer + grouped citations" [shape=box];

    "Frame the problem + stack" -> "Spawn one subagent per language (parallel)";
    "Spawn one subagent per language (parallel)" -> "Each: native terms -> WebSearch -> WebFetch strongest";
    "Each: native terms -> WebSearch -> WebFetch strongest" -> "Each: return structured findings";
    "Each: return structured findings" -> "Coordinator synthesis";
    "Coordinator synthesis" -> "Final answer + grouped citations";
}

Step 1 — Frame

Before spawning, write one short brief the subagents share:

• the problem / idea / question in one paragraph
• the exact stack (language, framework, package names, version numbers, error messages)
• what a good answer looks like (a fix, a list of approaches, a benchmark, etc.)

Step 2 — Spawn one subagent per language, IN PARALLEL

Launch all of these in a single message (multiple Agent tool calls) so they run concurrently. Use the general-purpose agent type (it has WebSearch + WebFetch and full tool access for open-web research):

• English Research Subagent
• Chinese Simplified Research Subagent
• Chinese Traditional Research Subagent
• French Research Subagent
• Russian Research Subagent
• Hebrew Research Subagent
• Spanish Research Subagent

Each subagent works independently in its own language and must:

1. Translate the problem into native technical search terms — not literal translation. Use the real terms a native engineer would type.
2. Build native-language WebSearch queries using: correct technical terms, synonyms, exact error messages, package names, framework names, version numbers, common workaround phrases.
3. WebSearch first.
4. WebFetch the strongest sources.
5. Search official sources, forums, GitHub, blogs, papers, benchmarks, and community discussions — not only official docs, but the professional forums where real-world solutions appear.
6. Return findings in the SAME structured format below — identical schema across all subagents, so the coordinator can compare directly.

Source priority (all languages)

1. official docs → 2. GitHub Issues/PRs → 3. GitHub Discussions → 4. Stack Overflow → 5. academic papers → 6. benchmark reports → 7. high-quality engineering blogs with reproducible details → 8. official vendor forums → 9. Reddit technical communities → 10. Hacker News → 11. Discord/Slack/community forums if publicly searchable → 12. language-specific developer forums.

Where each language subagent should look

Language	Native sources to include
English	GitHub Issues, GitHub Discussions, Stack Overflow, Reddit, Hacker News, official vendor forums, engineering blogs, project docs, package issue trackers
Chinese Simplified	CSDN, Zhihu, SegmentFault, Juejin, CNBlogs, Tencent Cloud dev articles, Alibaba Cloud dev articles, Chinese-language GitHub issues, Chinese tech blogs/forums
Chinese Traditional	iT 邦幫忙, Traditional-Chinese Medium posts, Taiwanese dev blogs, Taiwanese tech forums, Traditional-Chinese GitHub issues, TW vendor/community posts
Russian	Habr, Stack Overflow на русском, Russian-language GitHub issues, Russian dev blogs, Russian vendor/community forums, publicly indexed Telegram/forum mirrors
French	Developpez, Stack Overflow en français, French-language GitHub issues, French dev blogs, French vendor/community forums
Hebrew	Hebrew tech blogs, Israeli dev forums, Hebrew GitHub issues/discussions, Israeli tech community posts (if publicly indexed)
Spanish	Stack Overflow en español, Spanish-language GitHub issues, Spanish + Latin American dev blogs, Foros del Web, La Web del Programador, Comunidad Platzi, OpenWebinars, Spanish vendor/community forums, Spanish Reddit; Spanish YouTube only if it includes code/links/reproducible steps

Step 3 — Each subagent extracts (per finding)

• URL / title
• language
• source type: official docs / GitHub / forum / paper / blog / benchmark / vendor forum
• proposed solution
• evidence quality
• code examples or reproducible steps (if available)
• whether it applies to our exact stack
• date / freshness
• contradictions with other sources
• confidence level

Forum evaluation (apply when reading forums)

• Is the answer accepted or confirmed by many users?
• Did a maintainer or official representative respond?
• Is there working code?
• Is the solution recent?
• Does the workaround still apply to the current version?
• Are there warnings, edge cases, or reports it breaks something else?

Step 4 — Coordinator synthesis

After all subagents finish, do ONE synthesis. The coordinator must:

1. Compare findings across languages.
2. Identify repeated solutions that appear independently in multiple languages (strong signal).
3. Identify contradictions.
4. Prefer evidence from official docs, maintainers, reproducible GitHub issues, papers, and confirmed forum answers.
5. Downgrade weak forum-only claims unless confirmed elsewhere.
6. Explain which findings are language-specific and which are globally supported.

Final answer format

• Best solution
• Alternatives
• Risks
• What to test
• Sources grouped by language
• Citations grouped by language
• Clear note where evidence is strong vs weak
• Final confidence level

Notes

• Requires WebSearch + WebFetch (available in Claude Code). If the environment lacks web tools, tell the user and stop.
• Scale languages to the task: the 7 above are the default. Drop or add languages when the domain clearly favors specific ones, but keep at least English + 2 others — cross-language agreement is the whole point.
• Token-heavy by design (7 parallel subagents). Worth it for hard bugs, design ideas, and prior-art hunts; overkill for trivial lookups.