Google Search¶

Quick Reference¶

Overview¶

Use this skill to turn vague search requests into a disciplined search-and-verification workflow. Default to doing the search work for the user first, then give them the reusable Google queries and refinement strategy if they want to continue manually.

Mandatory Gates¶

Gates execute in strict serial order. Any gate failure blocks all subsequent steps.

1) Scope         2) Ambiguity     3) Evidence      4) Language
   Classification → Resolution   → Requirements  → Detection
   │                │               │               │
   category+goal    unclear?        what proof?     EN/CN/Both?
   → classify       → STOP+ASK     → define chain  → set strategy
        │                │               │               │
        5) Source       6) Mode         7) Budget       8) Execution
           Path       →    Selection  →    Control   →    Integrity
           │               │               │               │
           official first? Quick/Std/Deep  max queries     actually searched?
           → rank sources  → auto-select   → enforce       → report honestly

1) Scope Classification Gate¶

Map the request into one primary category and one goal before writing any query.

Categories: - Information: news, facts, latest status, company or person updates - Knowledge: tutorials, best practices, explainers, research, official docs - Materials: PDFs, reports, templates, images, datasets, downloads - Tools: apps, services, plugins, utilities, alternatives - Public-information lookup: public records, bios, publications, profiles - Programmer search: error debugging, API docs, code examples, benchmarks, RFCs

Goals: Know | Learn | Create | Complete a task

2) Ambiguity Resolution Gate¶

STOP and ASK only when the readings would send you to different sources, so that guessing wrong spends the whole budget on the wrong question: - The target entity is ambiguous (e.g., "苹果" — fruit or Apple Inc.?) - The request maps to categories with disjoint sources (e.g., "help me with Redis" — learn it, debug an error, or compare it with Memcached?)

For everything else — an unstated time scope, a goal that could be "know" or "learn" — pick the most likely reading, state it as an assumption in the first line of the conclusion, and search. A question costs the user a round trip; a stated assumption costs one sentence and is just as easy to correct.

3) Evidence Requirements Gate¶

Before writing any query, define the minimum evidence chain needed to support the conclusion at the expected confidence level. This determines what to search for, not just how to search.

Confidence is always one of exactly three values — High, Medium, Low. Never invent an intermediate label such as "Medium-High"; if you are between two levels, take the lower one.

How to use: After classifying the question (Gate 1) and resolving ambiguity (Gate 2), map it to one row above. The evidence chain sets the minimum sources you must find — do not write the conclusion until the chain is satisfied, or explicitly degrade (see Honest Degradation).

If the evidence chain cannot be satisfied after the query budget is exhausted, degrade to Partial or Blocked rather than presenting an unsupported conclusion.

4) Language Detection Gate¶

Determine the primary search language based on the evidence chain requirements: - English-first: global technology, open-source, RFCs, vendor docs, academic papers - Chinese-first: China-specific policy, domestic companies, local regulations - Both (paired queries): engineering best practices, production experience, mixed topics

Read references/chinese-search-ecosystem.md when Chinese sources are needed.

5) Source Path Gate¶

Choose the source path before exploring broad results. Prefer the source, not commentary about the source.

Default ranking: 1. Official site, official account, original publisher, original document 2. Primary data, paper, PDF, filing, release notes, standards body 3. Reputable media or institutions that cite original material correctly 4. High-quality specialist communities or vertical sites 5. Aggregators, reposts, SEO pages, and summaries

Use domain constraints early when the right source family is obvious.

Read references/source-evaluation.md for full ranking and conflict-resolution rules.

6) Execution Mode Gate¶

Auto-select mode based on task signals, then state the selection in output:

Mode definitions:

If the user explicitly requests a specific mode, use that mode.

Quick Mode Fast Path: When signals clearly point to Quick mode (simple factual question, single-answer expected), collapse gates 1–8 into a single-line internal check and skip gate execution log in the output. Do NOT output per-gate logs for Quick mode — go straight to queries and conclusion.

7) Budget Control Gate¶

Enforce bounded query budgets per mode: - Quick: max 2 queries - Standard: max 5 queries (Round 1: 3, Round 2: 2) - Deep: max 8 queries (Round 1: 3, Round 2: 3, Round 3: 2)

If the budget is exhausted without a satisfactory answer, stop searching and report what was found, what remains uncertain, and what next strategy would resolve the gap.

Read references/ai-search-and-termination.md for escalation and termination rules.

8) Execution Integrity Gate¶

Never claim a search was performed unless it actually ran. - If a query was not executed, do not present hypothetical results. - If a source was not opened, do not claim to have verified its content. - Never report "confirmed" when the evidence is only from snippets. - Distinguish between "I found X" and "search snippet mentions X." - Never report that a filter applied unless the results show it did. A site: query whose results come from other hosts ran unfiltered — see Execution Environment. - This skill is not a source. Its references tell you how to search; they are not evidence for the answer, including answers about search operators. Every mode, Quick included, runs at least one query, and a Full verdict rests on a page opened in this session. In the first live eval (2026-09-24) an agent answered "no live search needed, the skill's own reference answers it" and returned Quick · Full with no URL — a Full verdict with nothing a user could check.

Execution Environment — WebSearch Is Not Google¶

The reusable queries this skill hands back are Google syntax, for the user to run in Google. The queries you execute go through Claude Code's WebSearch tool, which is a different engine with different behaviour. Measured 2026-09-24 on Claude Code 2.1.281, one sample per probe — if results look different, re-probe rather than trust this table:

Rules that follow:

1. Prefer the tool's allowed_domains / blocked_domains parameters for domain constraints. They are enforced by the tool, and an unreachable domain fails loudly with a 400 instead of quietly returning other hosts.
2. Check that every filter held. After each executed query, look at the result hosts. If a site: or allowed_domains constraint was not honoured, the results are unfiltered: say so under Evidence chain status and do not treat them as coming from the site you asked for.
3. Date operators carry no weight in an executed query. WebSearch ignores them, so the date of every time-sensitive claim comes from the page you opened. Keep after: in the reusable queries — it does work in Google.
4. Stack Overflow is unreachable from here. Do not cite a Stack Overflow page you could not open, and never present a dev.to or Medium result returned for a site:stackoverflow.com query as Stack Overflow content. Go to the official docs, the project's GitHub issues, or the language forum instead, and hand the Stack Overflow query to the user under Reusable queries.
5. The operator support tiers in references/programmer-search-patterns.md describe Google, not WebSearch. The tool describes itself as US-only; Chinese queries still returned Chinese hosts in these probes, but Baidu-only and in-app content is not reachable either way.

Workflow¶

After passing all gates:

1. Build Query Sets — Prepare Primary, Precision, and Expansion variants (Quick mode may prepare only the first two, since it can execute at most two). Preparing a variant is not executing it: the Gate 7 budget caps executed queries only, and a prepared-but-unrun variant is reported as a reusable query marked not run. Read references/query-patterns.md for category-specific patterns. For programmer searches, read references/programmer-search-patterns.md.

2. Execute and Triage — Search with the strongest query first, then refine based on results. Open first-party or original sources before commentary. If the first pass is weak, reformulate by changing one variable at a time (see Refinement Loop in query-patterns.md).

3. Evaluate and Cross-Check — Treat results as candidate evidence, not truth. For each important source, judge: originality, recency, directness, specificity, independence. Read references/source-evaluation.md for full evaluation protocol.

4. Write the Answer — Follow the Output Contract below.

For high-conflict and high-change topics (wars, elections, disasters), read references/high-conflict-topics.md for stricter scope-locking and source-tiering.

Content Access Resilience¶

WebFetch may fail to extract content from sites behind Cloudflare, AWS WAF, or JavaScript-heavy SPAs. When this happens:

Failure Detection¶

Recognize blocked responses: - HTTP 403 Forbidden or empty body from a known-content page - Response contains "Just a moment...", "Checking your browser", "Enable JavaScript" - Extracted text is < 30 words from a page that should be content-rich

Fallback Chain¶

When WebFetch fails, try these in order:

1. Firecrawl — If the firecrawl-scrape skill is available, use it (handles JS rendering and anti-bot)
2. Snippet-only mode — Use search snippets as evidence, but explicitly label: "Based on search snippet, not full page content"
3. Platform-specific — Tell the user which platform to search directly (e.g., "This StackOverflow page requires browser access")

Reporting¶

When degraded to snippet-only, the answer must: - Set degradation level to Partial (not Full) - State which sources could not be fully accessed - Lower confidence labels accordingly - Provide the direct URL so the user can verify manually

Anti-Examples — DO NOT Do These¶

1. Presenting search snippets as verified facts — a snippet is a preview, not a confirmed source. Open the page and verify before citing.

   BAD: "According to search results, the answer is X."
   GOOD: "According to [specific source, date], the answer is X."
   

2. Running one vague query and declaring the search complete — prepare the Primary / Precision / Expansion set before searching. One executed query is enough only when it was the Precision variant and it landed an official or primary source that directly answers the question; anything less specific means running the next variant, not writing the conclusion.

3. Ignoring time-sensitivity — searching "latest Go version" with no anchor surfaces pages from 2019. Technical topics go stale fast.

   BAD: Go latest features
   GOOD: Go release history site:go.dev/doc/devel
   

   The release-history page lists every release with its date, so its newest entry answers "latest" without writing a version number into the query that will itself go stale. A date bound is not the fix: after: filters on when the document was last updated, not when the event happened, and WebSearch ignores it entirely (see Execution Environment). Take the date from the page itself.

4. Returning raw links without synthesis — the user wants an answer, not a link dump. Synthesize first, then provide sources and reusable queries.

5. Using Google for topics better served by platform-specific search — WeChat articles, Xiaohongshu reviews, and Douyin content are not well-indexed by Google. Tell the user which platform to search directly.

6. Verifying an AI-generated claim with another AI tool — use Google to find original sources. Do not verify AI with AI.

7. Searching indefinitely without stopping — respect the query budget. 8 queries without a satisfactory answer means the problem is framing, not insufficient searching.

8. Omitting confidence and source-tier labels on key numbers — every key numeric claim must carry both labels. If you cannot label it, do not present it as settled.

9. Assuming a filter worked because you typed it — site:stackoverflow.com in WebSearch returns dev.to and GitHub pages with no error. A result list you did not check against the filter is an unfiltered result list.

Honest Degradation¶

When search results are insufficient, degrade explicitly:

Decision tree:

Did the budget run out with no relevant evidence at all,
or is the content behind a paywall / inside a walled garden?
  → YES → Blocked (state what was not found + platform/tool recommendation)
  → NO  → Is every link of the Gate 3 evidence chain satisfied?
           → NO  → Partial (name the missing link; do not upgrade on the
                            strength of the links you did get)
           → YES → Do those sources answer the question that was asked,
                    directly, at the row's target confidence?
                     → YES → Full
                     → NO  → Partial (the chain is complete but the sources
                                      answer an adjacent question)

Ask about total failure first. Every Blocked case is also a case of an unsatisfied evidence chain, so testing the chain first would route "budget exhausted, nothing found" to Partial and leave Blocked unreachable — a Partial promises a qualified answer, which is precisely what you do not have.

The chain question comes before the source question because a strong source answering an adjacent question is the most common way a Partial gets mislabeled Full: Gate 3 asked for an official basis plus a practitioner report, one of the two is missing, and the official page that does exist is about something else. Standard and Deep chains have two or three links, so checking only the strongest source cannot tell you the chain is complete. And the final branch matters for the rows whose chain is not official-tier at all — a technology comparison is Full on two independent benchmarks with disclosed methodology, at target confidence Medium.

Never guess to fill gaps. State uncertainty clearly.

Safety Rules¶

1. For public-information lookups, use public sources only and separate fact from inference.
2. Never present identity linkage (two records = same person) as confirmed unless sources make it explicit.
3. When sources conflict, explain the conflict rather than picking a side silently.
4. For high-conflict topics, follow the full protocol in references/high-conflict-topics.md.
5. Every key numeric claim must carry both a confidence label and a source-tier label.
6. If the evidence is weak, say so and provide the next search strategy instead of guessing.
7. Public-information lookup covers a person's public role — publications, official positions, filings, press statements. Do not assemble a profile of a private individual (home address, phone number, family members, daily movements, or tying an anonymous handle to a real name), even when each piece is public on its own: the aggregation is the harm. Decline that part and say why.
8. Do not route the user to pirated copies. For books, courses, and paid reports, point to the publisher, a library, or the author's own page rather than to netdisk search engines.

Output Contract¶

Every completed use of this skill must include the fields below. Fields are graded MUST / SHOULD / MAY per mode:

Quick mode output shape: render fields 1–2 as a single leading line (Quick · Full), then the conclusion, then the key numbers with their labels, then the reusable queries. No tables, no per-field headings, no gate log.

What "keep it short" does not mean: fields 1, 2, 3, 7 and 8 are MUST in Quick mode and are never dropped for brevity. The whole of Quick mode is four short lines —

Quick · Full
<the answer, one or two sentences>
Key numbers: <value> (`High`, `Official`) · source: <url you opened>
Reusable queries: `<precision query>`, `<primary query>`

— so if the metadata is outgrowing the answer, the fix is to compress each field to one line, not to delete fields. An answer with no labels, no cited URL and no reusable queries has not used this skill; it has just answered the question.

Load References Selectively¶

For every search task, before building queries: → Load references/query-patterns.md for core query construction patterns (Primary / Precision / Expansion variants), category-specific operators, and one-variable-at-a-time refinement strategy.

When the search category is Programmer search (error debugging, API docs, code examples, benchmarks, RFC lookup): → Load references/programmer-search-patterns.md for error message normalization, site: operator shortcuts per source type (GitHub, Stack Overflow, official docs), and technical evidence quality criteria.

When evaluating whether sources are strong enough to support a claim, or when sources conflict: → Load references/source-evaluation.md for the source ranking table (Official > Academic > Authoritative secondary > …), conflict resolution rules, originality / recency / directness / independence scoring criteria.

When the topic is Chinese-language or China-centric (content lives in Baidu, Zhihu, WeChat, Weibo): → Load references/chinese-search-ecosystem.md for content-type-to-source mapping, Chinese-specific search operators, and cross-validation rules for Chinese-only sources.

When the topic involves active conflict, elections, disasters, market moves, or any fast-changing event: → Load references/high-conflict-topics.md for stricter scope-locking rules, source-tiering requirements (no single outlet, require independent confirmation), and temporal anchoring conventions.

When assessing whether to stop searching or escalate to a different tool (AI search, database query): → Load references/ai-search-and-termination.md for budget exhaustion thresholds, escalation signals (no new evidence after N queries), and AI-search-vs-Google decision criteria.

When calibrating output quality or validating the report format for the first time: → Load references/worked-examples.md for fully worked examples across Quick / Standard / Deep modes with complete output contract fields, evidence chain notation, and reusable query formatting.

Worked Examples (Skeleton)¶

Full worked examples with complete Output Contract fields: read references/worked-examples.md.

Example 1: Quick mode¶

"sync.Pool GC 回收?" → Gates pass → Evidence chain: 1 official source → 2 executed queries (site:go.dev) → Full degradation → answer with High + Official labels, the URL of the page actually opened, and 2 reusable queries. A Full verdict without a citable URL is not Full.

Example 2: Standard mode¶

"MySQL 连接池配置" → Gates pass → Evidence chain: 1 official basis + 1 practitioner report → 4/5 executed queries (EN + CN) → Partial degradation, because the vendors document what the knobs do and no source gives an official sizing number → documented defaults labeled High + Official, the community headroom factor labeled Low + Practitioner report → gate execution log.