Conducting deep research

One search does not settle a hard question. This skill is the method for the harder kind: take a question, work it from several angles, read the sources instead of the snippets, cross-check what they say, and hand back a report that cites where every claim came from and is honest about what is still unknown.

It runs on Scout, which gives you live web search, full page reading, and structured extraction through one API. The steps below are tool-agnostic. See Reaching Scout for the two ways to call it.

When to use this

Use it when the answer needs more than a single lookup:

• "Research X and write me a briefing."
• "Is claim Y actually true? Find the primary source."
• "Compare A, B, and C on price, limits, and track record."
• "What happened with Z, and who reported it first?"
• "Map the current state of <market / technology / debate>."

For a single fact you already know how to look up, just search once. This method is for questions where the work is in the cross-checking.

The method

Run these in order. Do not skip to writing before you have read sources.

1. Frame the question. Restate it in your own words. Break it into the sub-questions you would need to answer to call it done. Decide what a good answer looks like before you start, so you know when to stop.
2. Plan the searches. Each sub-question usually needs more than one query, from more than one angle. List the angles first. A claim and its rebuttal often live under different search terms, so plan to look for both.
3. Sweep broad. Run the searches and collect candidate sources. Keep the ones worth reading; note why. For a question that resists a flat search, use Scout's deep search, which works the topic across many queries and streams progress. See references/searching.md.
4. Read the sources. Open the promising ones and read the actual page, not the search snippet. Snippets are written to get a click and routinely drop the qualifier that changes the meaning. Scrape the page or extract the specific fields you need. See references/sources.md.
5. Cross-check. A claim is only as good as its sourcing. Corroborate each load-bearing claim against an independent source. When sources disagree, that disagreement is a finding, not a problem to hide. Track who said what and when.
6. Synthesize. Write the report. Cite every claim inline. Separate what the sources support from what you are inferring. State your confidence and list what you could not pin down. See references/report.md.

The loop is not always clean. Reading sources raises new sub-questions; go back to step 2 and search again. Stop when new searches stop changing the answer, not when you run out of patience.

Reaching Scout

Get an API key at https://platform.usescout.sh/settings and put it in the SCOUT_API_KEY environment variable.

If the Scout MCP tools are available (you will see tools named scout_search, scout_scrape, scout_extract, scout_company, scout_answer, scout_find_all), call those directly. They are the cleanest path and handle auth for you.

Otherwise call the REST API at https://core.usescout.sh:

# Search the live web
curl -s https://core.usescout.sh/v1/search \
  -H "Authorization: Bearer $SCOUT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "your query", "num_results": 10}'

# Read a page as clean markdown
curl -s https://core.usescout.sh/v1/scrape \
  -H "Authorization: Bearer $SCOUT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com/article"}'

# Pull specific fields out of a page
curl -s https://core.usescout.sh/v1/extract \
  -H "Authorization: Bearer $SCOUT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com", "schema": {"price": "string", "released": "string"}}'

The reference files show when to reach for deep search, how to judge a source, and how to write the report. Load them as you hit each step. Do not paste them into your context up front; read each one when you get to its step.

What good looks like

• Every factual claim has a source next to it.
• Load-bearing claims are corroborated by a second, independent source.
• Disagreement between sources is reported, not averaged away.
• The report says what is well-supported and what is thin.
• A reader can follow your citations and reach the same conclusion.