internet-search

name	internet-search
description	Tool-agnostic search — query construction, tool selection, source trust hierarchy.
context	fork
globs	["**/*.md"]
alwaysApply	false

Internet Search Skill

1. Tool Selection

Three search tools are available. Choose by task type:

Task	Primary Tool	Why	Fallback
Library/framework docs, API reference, version migration	context7	Official docs + code examples, version-aware	webfetch known docs URL
Broad web search, technical Q&A, current info, news	websearch	Semantic search across indexed web, live crawl option	webfetch specific result URLs
Known URL (docs page, GitHub issue, blog post, changelog)	webfetch	Direct fetch of a specific page, returns markdown	—

Parallel rule: context7 + websearch can run simultaneously — they are independent tools.

---

2. Query Construction

2.1 Be Specific

❌ "Python async"
✅ "FastAPI background tasks SQLAlchemy async session 2024"

❌ "Docker error"
✅ "Docker multi-stage build COPY --from permission denied non-root"

2.2 Include Version

Append the target version or year to scope results:

"SQLAlchemy 2.0 async session pattern"
"Next.js 15 server components migration"
"React 19 use() hook"

2.3 Use Domain-Specific Keywords

# Framework/library — include name + key concept
"FastAPI dependency injection lifespan"

# Bug/error — include exact error message + language/framework
"TypeError: 'NoneType' object is not callable FastAPI"

# Comparison — make explicit
"Redis vs Valkey 2025 performance comparison"

---

3. Source Trust Hierarchy

Always escalate from the highest trust tier before falling back:

Tier	Sources	Use for
T1 — Authoritative	Official vendor docs, RFC/specs, peer-reviewed papers, GitHub READMEs	Definitive answers, API contracts, algorithm correctness
T2 — High-signal	Stack Overflow (accepted/high-vote), GitHub Issues/Discussions, official changelogs	Real-world patterns, known bugs, version behaviors
T3 — Useful, verify	Hacker News, Reddit (engineering subs), well-cited blogs	Current opinions, ecosystem trends, practical tips
T4 — Last resort	Random tutorials, Medium articles, LLM-generated content	Background only — always cross-check with T1/T2

Never cite T3/T4 as the sole source for a technical decision. Always pair with a T1 primary source.

---

4. Parallel Search Strategy

STEP 1 — PARALLEL (T1 sources):
  ├── context7 (library docs — structured, version-aware)
  └── websearch (broad web + community)

STEP 2 — webfetch (T1/T2 specific URLs):
  ├── Known docs page or GitHub README
  ├── Stack Overflow accepted answer
  └── GitHub issue / discussion

STEP 3 — ONLY IF NEEDED (T3):
  └── websearch with site:reddit.com or site:news.ycombinator.com

→ Synthesize after each step; stop when enough evidence gathered.

For known URLs, use webfetch directly. For search results that return URLs, webfetch the most promising ones for full content.

---

5. Result Format

Always return structured findings:

## Research: {TOPIC}

**Answer**: [Direct answer, 1-2 sentences]

**Confidence**: High / Medium / Low
**Sources**:
| URL | Tier | Key Takeaway |
|-----|------|-------------|
| [official docs](url) | T1 | Confirms pattern X |
| [SO answer](url) | T2 | 342 votes — accepted fix |
| [GitHub issue #123](url) | T2 | Fixed in v2.1.0 |

**Caveats**: [Version-specific warnings, conflicting evidence, stale info]

---

6. Quality Checklist

Before finalizing, verify:

• Recency — info still valid? Flag anything >3 years old in fast-moving fields (JS/TS, LLMs, cloud infra).
• Version alignment — does the answer match the project's library/major version?
• Primary over secondary — prefer official docs over tutorials that re-explain them.
• Confidence calibrated — T3/T4 sources downgrade confidence. Single source = Low.
• No hallucinated URLs — only cite URLs from actual fetch results or officially documented patterns.