Deep Research

What this skill is for

deep-research performs live web research that goes beyond a single URL fetch — searching deliberately across multiple sources, identifying the right evidence, testing the first conclusion, and returning a recommendation that separates observed facts from judgment. It is for research work that would be weak if handled as a few quick page fetches.

This skill is normally called internally by 100-domain-knowledge for non-trivial domain research, but it can be used directly any time a well-sourced, current-evidence answer is needed.

When to use it

When the answer to a question would change based on current, real-world evidence
For vendor comparisons, tool evaluations, framework comparisons, and technology decisions
For market scans, industry analysis, competitive research, and due diligence
When the question includes words like “latest”, “current”, “what changed”, or “right now”
When 100-domain-knowledge needs live research as its primary research mechanism

When NOT to use it

For summarizing a single URL the user has already provided — that is a webfetch task
For simple fact lookups that do not require source discovery and synthesis
When the question is really a code change, document edit, or data processing task with minimal web research involved
When model memory is sufficient and freshness is not a concern

What it produces

A synthesized research report with:
- Key findings organized by research lane (official docs, vendor sources, independent analysis, community evidence)
- Citations to primary sources with URLs
- Explicit distinctions between observed facts, inferred implications, and the recommendation
- Trade-offs and uncertainty made explicit
- At least one plausible alternative tested before the recommendation is made

Required capabilities

deep-research requires live web access. In OpenCode, the default discovery path is Brave Search result pages (https://search.brave.com/search?q=<query>) for primary search, with Bing result pages (https://www.bing.com/search?q=<query>) as secondary fallback, combined with webfetch for page extraction. Either engine may still return anti-bot or access-control pages to automated requests, so the skill should detect that outcome and report discovery as blocked for that engine rather than pretending results were retrieved. If neither search nor page retrieval is available, the skill states this rather than presenting model memory as live research.

In practice, the current stack works like this:

Brave first for initial source discovery
Bing second when Brave is blocked, empty, or not usable from the current environment
webfetch last to read the actual source pages once promising URLs have been identified

This means deep-research is currently using live search result pages for discovery, not a dedicated search API. The goal is to find relevant source URLs first, then fetch and synthesize the underlying pages.

Common prompts

Prompt

Run deep-research on the current state of HIPAA compliance requirements for cloud-hosted SaaS applications.

Prompt

Run deep-research — compare SWR, TanStack Query, and Apollo Client for a React app with a REST API.

Prompt

Use deep-research to find what changed in the EU AI Act since 2024 that affects automated decision systems.

Research approach

deep-research follows a structured workflow — not a single query:

Frame the question — pin down the actual decision, extract constraints, write a research brief
Plan search lanes — 3–6 distinct search angles (official docs, vendor sources, independent analysis, news, community evidence)
Discover sources — run searches, evaluate source quality, find additional leads
Extract evidence — fetch relevant pages, extract specific claims with citations
Challenge the first conclusion — test at least one plausible alternative before finalizing
Synthesize — separate facts, inferences, and recommendation; make uncertainty explicit

Practical guidance

The research brief is the most important input. A vague question produces a broad survey. A specific question produces actionable evidence.
Primary sources come first: official documentation, maintainer statements, standards bodies, government publications. Secondary sources add context.
Search result pages are only the discovery layer. The actual evidence should come from the destination pages that are fetched afterward.
Important claims should be verified across multiple sources where possible — not accepted from a single page.
Do not treat model background knowledge as live research for freshness-sensitive questions. If the question involves “current”, “latest”, or “recent”, live discovery is required.

Common mistakes

Asking for research on a vague topic and expecting a specific recommendation
Treating a single-URL summary as deep research
Accepting the first conclusion without testing an alternative
Using model memory for questions that require current external evidence