Web Research Subagent

Use this skill to keep the main context clean: the subagent explores the web, then returns only decision-useful facts with citations.

For cross-client usage (Cursor/Claude Code/OpenCode), read:

• references/client-adapters.md
• references/AGENTS.template.md

Workflow

1. Build a narrow research brief

Before delegating, define:

• Goal: what decision the main agent needs to make
• Scope: allowed domains, date range, locale, and required recency
• Output contract: max findings, citation format, and uncertainty notes

Keep the brief under 12 lines.

2. Delegate research to a subagent

Spawn one subagent dedicated to web research. Give it:

• The brief from Step 1
• A strict output format (see references/subagent_prompt_template.md)
• A requirement to provide source URLs for each claim

Do not pass unrelated implementation context.

3. Enforce context hygiene on return

When the subagent reports back:

• Keep only high-signal findings (usually 3-7 bullets)
• Drop repetitive details and generic background
• Preserve links, dates, and confidence notes
• Mark inference vs. source-stated facts

Never paste full search dumps into main context.

4. Integrate for the main task

Convert the research result into the minimum needed for action:

• Decision summary (what matters now)
• Evidence table (claim -> source link -> date)
• Known unknowns (what is still uncertain)

If confidence is low, run one targeted follow-up query instead of broad re-search.

Output Contract (Required)

Require this exact structure from the subagent:

1. Answer
2. Key Findings (3-7 bullets)
3. Evidence (claim + URL + publish/update date)
4. Assumptions and Gaps
5. Confidence (high/medium/low)

Reject outputs that do not include URLs.

Guardrails

• Prefer primary sources (official docs, standards, vendor pages, first-party announcements).
• Use secondary sources only when primary sources are missing, and label them.
• Quote sparingly; summarize in your own words.
• If multiple sources conflict, show both and explain which is more reliable.

Quick Trigger Examples

• "Use this skill to research the latest API pricing and return only citations I need."
• "Delegate web search to a subagent; give me concise facts and links, no raw dumps."
• "Find current policy changes online and report only actionable deltas."