deep-research

Deep Research

Drift-prone skill. Exa/web MCP tool names, quotas, and result shapes change. Verify the configured MCP tools and current API docs before promising coverage or quoting live source counts.

Produce thorough, cited research reports from multiple web sources, and route lighter asks to the cheapest path that answers them. This is the single research entry point — it absorbs the "classify the ask, take the lightest useful path first" routing and the evidence-labeling discipline that used to live in a separate research-ops skill.

When to Activate

• User asks to research a topic in depth, compare options, or evaluate a decision
• Competitive analysis, technology evaluation, market sizing, or due diligence
• Any question requiring synthesis from multiple current sources
• User says "research", "deep dive", "compare", or "what's the current state of"

Do not spin up a full pass when the answer is already in local code/docs, or when a single quick lookup suffices — see Step 0.

MCP Requirements

This repo's configured Exa server exposes:

• web_search_exa — web/news discovery
• web_fetch_exa — fetch a URL's full content (use on the best result URLs when snippets aren't enough, including code/docs pages)

Any configured search/fetch MCP works as a substitute — verify the exposed tool names first (the exa-search skill documents the current Exa surface). Configure in ~/.claude.json or ~/.codex/config.toml.

Workflow

Step 0: Classify the ask (take the lightest path first)

Before searching, pick the lane:

• Quick factual answer → one web_search_exa (or the exa-search skill); stop there.
• Comparison / decision memo → the full multi-source pass below.
• Already in local repo/docs → answer from there; don't run a heavy pass.
• Recurring question → answer now, then say so and recommend a monitor/workflow layer instead of repeating the same manual search forever.

Also normalize anything the user already supplied into: already-evidenced facts · needs verification · open questions. Don't restart from zero if the user already built part of the model.

Step 1: Understand the Goal

Ask 1-2 quick clarifying questions:

• "What's your goal — learning, making a decision, or writing something?"
• "Any specific angle or depth you want?"

If the user says "just research it" — skip ahead with reasonable defaults.

Step 2: Plan the Research

Break the topic into 3-5 research sub-questions. Example:

• Topic: "Impact of AI on healthcare"
  • What are the main AI applications in healthcare today?
  • What clinical outcomes have been measured?
  • What are the regulatory challenges?
  • What companies are leading this space?
  • What's the market size and growth trajectory?

Step 3: Execute Multi-Source Search

For EACH sub-question, search using the configured tools:

web_search_exa(query: "<sub-question keywords>", numResults: 8)

Search strategy:

• Use 2-3 different keyword variations per sub-question
• Mix general and news-focused queries
• Aim for 15-30 unique sources total
• Prioritize: academic, official, reputable news > blogs > forums

Step 4: Deep-Read Key Sources

For the most promising URLs, fetch full content — don't rely on snippets alone:

web_fetch_exa(url: "<url>")   # for any URL, including GitHub/SO/docs pages for code & API detail

Read 3-5 key sources in full for depth.

Step 5: Synthesize and Write Report

Structure the report:

# [Topic]: Research Report
*Generated: [date] | Sources: [N] | Confidence: [High/Medium/Low]*

## Executive Summary
[3-5 sentence overview of key findings]

## 1. [First Major Theme]
[Findings with inline citations]
- Key point ([Source Name](url))
- Supporting data ([Source Name](url))

## 2. [Second Major Theme]
...

## Key Takeaways
- [Actionable insight 1]
- [Actionable insight 2]

## Sources
1. [Title](url) — [one-line summary]
2. ...

## Methodology
Searched [N] queries across web and news. Analyzed [M] sources.
Sub-questions investigated: [list]

Label every important claim by evidence type — sourced fact vs user-supplied context vs inference vs recommendation — and give concrete dates on freshness-sensitive answers.

Step 6: Deliver

• Short topics: Post the full report in chat
• Long reports: Post the executive summary + key takeaways, save full report to a file

Parallel Research with Subagents

For broad topics, use Claude Code's Task tool to parallelize:

Launch 3 research agents in parallel:
1. Agent 1: Research sub-questions 1-2
2. Agent 2: Research sub-questions 3-4
3. Agent 3: Research sub-question 5 + cross-cutting themes

Each agent searches, reads sources, and returns findings. The main session synthesizes into the final report.

Quality Rules

1. Every claim needs a source. No unsourced assertions.
2. Cross-reference. If only one source says it, flag it as unverified.
3. Recency matters. Prefer sources from the last 12 months; date freshness-sensitive claims.
4. Label evidence type. Separate sourced fact, user-supplied context, inference, and recommendation.
5. Acknowledge gaps. If you couldn't find good info on a sub-question, say so.
6. No hallucination. If you don't know, say "insufficient data found."

Examples

"Research the current state of nuclear fusion energy"
"Deep dive into Rust vs Go for backend services in 2026"
"Research the best strategies for bootstrapping a SaaS business"
"What's happening with the US housing market right now?"
"Investigate the competitive landscape for AI code editors"