web-search Skill

When to Use

ALWAYS invoke this skill for any task involving:

Web search
Research or deep investigation
Scraping or data extraction from websites
Finding latest/current information
Comparing options from web sources
Any "find on the web" or "research X" request

Never perform web research using only transient memory or single-channel tools.

Core Principles (MUST Follow)

Browser-First Anti-Bot Strategy: Prioritize human-like browser scraping using
```
agent-browser
```
(or
```
playwright-cli
```
via browser-switch) for rendered pages to avoid blocks.
Parallel Retrieval: Launch multiple subagents/workers in parallel (Task tool when available), each with a distinct role:
- Discovery: Prefer
```
searxng-search
```
  /
```
searxng-extract
```
  when installed; otherwise native web search, web fetch, Tavily, exa, or other available discovery tools.
- Scraping:
```
agent-browser
```
  (human-like navigation + snapshots);
```
human-search
```
  cascade as fallback.
- Synthesis:
```
deep-research
```
  patterns toward
```
90_synthesis.md
```
  .
- One topic per research subagent: When the user asks about multiple distinct subjects, spawn separate subagents so each worker owns one topic only — each with its own
```
docs/research/YYYYMMDD/{research_topic_slug}/
```
  folder. Do not assign unrelated questions to the same research subagent (avoids cross-contaminated context, checkpoints, and citations).
Immediate Persistence: After every single source (search result, URL extracted, page rendered), append a structured checkpoint to the numbered artifact for that phase (e.g.
```
10_discovery.md
```
,
```
20_sources.md
```
). Do not proceed to the next source until the checkpoint is written. Required fields, example block, and persistence edge cases →
```
references/checkpoint-template.md
```
.
Dated Artifact Convention: Use
```
docs/research/YYYYMMDD/{research_topic_slug}/
```
— for example
```
docs/research/20260430/ai_coding_agents/
```
. The research topic is carried by
```
{research_topic_slug}
```
, not baked into repeated strings inside filenames. Under that folder reuse the same phase filenames every time so layouts stay predictable. Typical files:
- ```
00_plan.md
```
- ```
10_discovery.md
```
- ```
20_sources.md
```
  (checkpoints for fetched pages: browser, extract, etc.)
- ```
90_synthesis.md
```
- ```
YYYYMMDD
```
  : Windows (PowerShell) →
```
Get-Date -Format "yyyyMMdd"
```
  ; Linux/macOS (shell) →
```
date +%Y%m%d
```
  ; generic (any OS) → Python
```
datetime.date.today().strftime("%Y%m%d")
```
  .
Synthesis from Disk Only: The final answer and
```
90_synthesis.md
```
must derive exclusively from persisted artifacts, never solely from conversation memory.
Citations and Traceability: Follow
```
deep-research
```
patterns; citations must tie back to checkpoint entries and on-disk files.

Workflow checklist

Operational order for a research run. Tool install and orchestration roles → See also — referenced skills; checkpoint layout and fields →

references/checkpoint-template.md

Receive research query
Partition distinct topics → one
```
{research_topic_slug}
```
folder each under
```
docs/research/YYYYMMDD/
```
Launch parallel subagents (Task tool): one subagent per topic for research workers; within each topic, parallelize by role (discovery / scraping / synthesis) as needed
Instruct each parallel subagent to use the referenced web-scraping and research skills and to save dated research files and checkpoints after each finding to avoid losing context. Each subagent should follow this skill’s rules.
After each source fetch: immediately append checkpoint
On write failure: retry once then use fallback file
After sufficient sources: run synthesis using deep-research patterns
Produce final response with citations linking back to persisted files
Confirm every expected artifact exists on disk and treat the run as incomplete until both synthesis and the final answer are grounded in those files (not conversation memory alone)

See also — referenced skills (install + orchestration roles)

Install missing skills by copying each repo’s skill folder into your agent’s skills directory.

web-search

orchestrates these tools rather than replacing them:

human-search: Primary intelligent cascade. Tiered fallback (native → Python scraper → browser CLI → crawl4ai); use as the default retrieval engine for robustness.
agent-browser: Primary anti-bot scraping. Workflow:
```
open
```
→
```
snapshot -i
```
→ extract using refs →
```
re-snapshot
```
after changes. Use named sessions for parallel work. Strongest human-like resilience.
searxng-search + searxng-extract: Fast, free, unlimited local discovery and extraction (no API keys). Ideal for gathering initial URL candidates.
deep-research: Final synthesis phase — structured reporting, citation management, progressive disclosure, professional formatting for
```
90_synthesis.md
```
.
browser-switch: Picks agent-browser, playwright-cli, or other browser backends based on context.
playwright-cli: Backup browser automation when agent-browser is unavailable or when specific Playwright features are needed.

Tip: After installing missing skills, tell your subagents the skill paths to use, otherwise they might not discover the skills until the agent restart.

web-search

NPX Install

Tags

SKILL.md Content

web-search Skill

When to Use

Core Principles (MUST Follow)

Workflow checklist

See also — referenced skills (install + orchestration roles)