Research

Research: Multi-Source Synthesis

What it does

Runs parallel research across several angles of a topic, each in its own subagent, then triangulates the findings and writes a cited report. Two modes: Quick (3 directions, executes immediately) for narrow questions, and Deep (up to 5 directions, posts a plan and waits for approval) for broad or comparative topics.

Requirements

• Web search and fetch tools in the agent runtime. Free option: Claude Code's built-in WebSearch and WebFetch.
• Subagents for parallel directions. Free option: Claude Code's Agent tool; without it, run directions sequentially.

Inputs and outputs

Worked example

Paste into Claude Code with this skill installed:

/agent-ops:research deep research: state of MCP server hosting options in 2026, compare managed vs self-hosted, freshness 6mo

Expected: a plan with 4-5 directions is posted and waits for "go"; after approval the report lands at ./research/mcp-server-hosting-20260907/report.md with one raw/direction-N.md per direction and inline citations.

Procedure

Deep triggers: user says "deep research", "deep dive", "comprehensive", or topic is broad/multi-faceted. Everything else defaults to Quick.

1. Scope

Determine mode, domain (tech/business/academic/general), freshness (7d/6mo/12mo/2-3yr/none), and restate the core question.

Read modules/{domain}.md and references/source-tiers.md.

If the plan depends on Reddit or forum evidence, also read references/reddit.md before retrieval. It defines the supported access path, sampling rules, and how to qualify community evidence.

2. Plan

Design broad, multi-angle directions. Each subagent has its own context window; make directions meaty so each covers 2-3 related subtopics.

Good: "Customer perception & brand positioning: reviews from multiple sites, website messaging, pricing structure, CRO changes via Wayback Machine." Bad: "Reviews analysis" (wastes an agent slot on a narrow angle)

Quick: 3 directions. Deep: up to 5 directions. If the user specifies a number, use that instead.

Plan format:

Research Plan: {topic}
Mode: Deep | Domain: {domain} | Freshness: {freshness}

Directions:
1. {direction 1: with scope description}
2. {direction 2}
...

Reply "go" to start, or suggest changes.

After sending the plan, you are done for this turn. Make no further tool calls. The user's next message will arrive as a new turn. Only proceed to step 3 when that message approves the plan.

Quick mode: skip approval, go directly to step 3.

3. Retrieve

Run date +%Y-%m-%d for today's date.

Step 2: Launch all directions as parallel subagents in a SINGLE message. Each call runs concurrently with its own context window.

For each direction, use the Agent tool (Claude Code) or equivalent Task/subagent primitive, with a general-purpose worker profile:

Agent({
  description: "Research direction N",
  prompt: `You are a research agent.

Direction: {direction text}
Topic: {overall topic}
Today: {YYYY-MM-DD}
Freshness: {constraint}
Output file: ./research/{slug}/raw/direction-N.md

- Use WebSearch extensively (10-20+ searches). Go deep.
- Use WebFetch to read full articles from quality sources. Try alternatives if a site blocks you.
- Write incrementally: after every 3-4 searches, update the output file with accumulated findings.
  First write creates the file. Subsequent writes replace it with the full report so far.
- Structure as markdown: key findings with source URLs, confidence levels (high/medium/low), contradictions.
- Keep under 5000 words. Summarize rather than dump raw data.
- When done, confirm the file was written and its approximate word count.`,
  model: "sonnet"
})

CRITICAL: Launch ALL directions in a single message; do NOT wait for one to finish before launching the next. Parallel subagents run concurrently.

Step 3: Verify:

wc -c ./research/{slug}/raw/direction-*.md 2>/dev/null

If any file is missing or under 500 bytes, retry those directions once with a single new Agent call. If still missing, note the gap and proceed with available data.

4. Synthesize

Read all raw/direction-N.md files. Cross-reference findings:

• Same claim from 2+ sources → state as fact with citations
• Single-source claim → qualify: "According to [Source], ..."
• Contradictions → present both sides with their sources
• Deduplicate URLs across directions

Write the report with [N] numbered inline citations. Include a bibliography at the end:

[1] Title - URL (Tier, accessed YYYY-MM-DD)

Include a Limitations section noting gaps, failed directions, and unverifiable claims.

Deep mode: save as report.md in the research folder. Quick mode: draft inline (no file needed).

5. Deliver

Read back the report. Verify: claims have citations, no unsourced facts, contradictions handled, limitations noted. Fix issues before delivering.

Quick mode: Send the full report (under 4000 chars). Split into two messages if longer.

Deep mode: Send a summary with:

• Key findings (5-8 bullets with source links)
• Note that the full report is saved in the research folder
• Under 4000 chars

Read references/source-tiers.md for the full tier system.

for i in 1 2 3 4 5; do
  [ -f /tmp/research_prompt_$i.txt ] || continue
  env -u CLAUDECODE timeout 600 claude -p "$(cat /tmp/research_prompt_$i.txt)" \
    --tools "WebSearch,WebFetch,Read,Write,Bash" \
    --strict-mcp-config \
    --disable-slash-commands \
    --effort low \
    --dangerously-skip-permissions \
    --no-session-persistence \
    --fallback-model haiku \
    --verbose \
    --model sonnet \
    --max-turns 40 > /tmp/research_log_$i.txt 2>&1 &
  echo "Launched direction $i (PID $!)"
  sleep 5
done
wait

Write prompts to /tmp/research_prompt_N.txt first (same format as the Agent prompt above). Set the Bash timeout to 600000ms.