mweb-research

name: "m:web-research" description: Web search, research, and page scraping via Perplexity + Firecrawl argument-hint: "" allowed-tools: [Bash, Read]

Web Research

You have been given the following research request:

$ARGUMENTS

Your Task

Analyze the request, choose the appropriate mode, and run the command. Then synthesize results for the user.

Step 1: Choose mode

If it's a URL (starts with http/https) → use --scrape
If it's a simple factual question (quick lookup, single fact) → use --ask
If it's a comparison or multi-source analysis → use --research
If it needs reasoning about tradeoffs or decisions → use --reason
If it needs exhaustive multi-source deep coverage AND the user explicitly asks for deep research → use --deep
If the user explicitly wants raw ranked search results → use --search

Default to --research when unsure. Do NOT use --deep unless the user explicitly asks for exhaustive/deep coverage (it's the slowest and most expensive mode).

Step 2: Run the command

uv run --directory ~/.claude python scripts/web_research.py --<mode> "$ARGUMENTS"

IMPORTANT: Always use --directory ~/.claude so uv finds the pyproject.toml with aiohttp dependency.

Step 3: Follow up (bounded)

If the initial results include source URLs that look highly relevant and the answer feels incomplete, follow up by scraping up to 4 of the most promising URLs. Run scrapes in parallel where possible.

Budget: Maximum 10 research/ask/reason calls, with up to 4 scrapes per research call to read the most relevant sources in full. Synthesize what you have once you've exhausted useful leads or hit the limit.

Step 4: Synthesize

After gathering results, write a clear summary for the user. Include source URLs at the end. Do NOT make additional API calls after synthesizing.

Command Reference

Modes

Mode	Perplexity Model	Purpose	Cost
`--ask`	sonar	Quick fact with AI answer	Low
`--search`	(direct search)	Ranked results, no AI synthesis	Low
`--research`	sonar-pro	AI combines multiple sources	Medium
`--reason`	sonar-reasoning-pro	Chain-of-thought analysis	High
`--deep`	sonar-deep-research	Exhaustive multi-source research	Very High
`--scrape`	(firecrawl)	Fetch full page content from URL	Low

Search options (for `--search`)

Parameter	Description
`--max-results N`	Number of results (1-20, default: 10)
`--recency`	Filter: `day`, `week`, `month`, `year`
`--domains`	Limit to specific domains

Scrape options (for `--scrape`)

IMPORTANT: --scrape takes ONLY a URL. Do NOT pass additional query text after the URL -- it will cause an argparse error.

# Correct:
uv run --directory ~/.claude python scripts/web_research.py --scrape "https://example.com/page"

# WRONG (will fail):
uv run --directory ~/.claude python scripts/web_research.py --scrape "https://example.com/page" "tell me about X"

Parameter	Description
`--format`	`markdown`, `html`, `text` (default: markdown)
`--full-page`	Include sidebars/nav (default: main content only)

Global options

Parameter	Description
`--model`	Override Perplexity model selection

API Keys Required

PERPLEXITY_API_KEY — for search/ask/research/reason/deep modes
FIRECRAWL_API_KEY — for scrape mode

Both loaded from environment or ~/.claude/.env.