Agent-Skills.md

Agent Skills: Research

Use when researching a topic via multi-source synthesis (literature review, competitive analysis, tech comparison); for a fact-checked cited web report use deep-research. Triggers on "research this", "investigate", "survey the landscape".

UncategorizedID: buvis/home/research

Repository

buvis/home
buvis
2

Install this agent skill to your local

pnpm dlx add-skill https://github.com/buvis/home/tree/HEAD/.claude/skills/research

Skill Files

Browse the full folder contents for research.

Download Skill

Loading file tree…

.claude/skills/research/SKILL.md

Skill Metadata

Name
research
Description
Use when researching a topic via multi-source synthesis (literature review, competitive analysis, tech comparison); for a fact-checked cited web report use deep-research. Triggers on "research this", "investigate", "survey the landscape".

Research

Dispatch a subagent to do multi-source research. Keeps intermediate search noise out of the main context window.

When This Skill Triggers

  1. Clarify intent if the request is vague (1-2 questions max). Unattended (CLAUDE_UNATTENDED=1): skip questions — research the request as stated and note assumptions in the report, per ~/.claude/skills/run-autopilot/references/unattended-contract.md.
  2. Launch a general-purpose Agent (model: sonnet) with the prompt template below. (sonnet is a deliberate cost pin: research is web search + synthesis, not deep reasoning — Sonnet handles the fuzzy-boundary summarization well at a fraction of Opus's cost, and the subagent already isolates the search noise. Override to opus only for a genuinely hard synthesis.)
  3. Report the synthesized result to the user.

For broad topics with 3+ independent angles, dispatch multiple agents in parallel (one per sub-question), then synthesize their results yourself.

Agent Prompt Template

Adapt this to the specific research question. Replace [TOPIC], [SUB_QUESTIONS], and [DEPTH].

You are a research agent. Your job is to thoroughly research a topic and return a structured report. Use WebSearch and WebFetch as your tools.

## Research question
[TOPIC]

## Sub-questions to investigate
[SUB_QUESTIONS - 3-5 focused angles]

## Depth
[DEPTH - "quick survey" or "exhaustive review"]

## Process

1. Run 15-30 WebSearch queries across the sub-questions. For each, vary phrasing:
   - Direct question form
   - Keyword-dense form
   - Site-scoped where useful (e.g. arxiv.org, official docs)
   Prioritize: academic papers, official docs, project repos, reputable technical blogs.
   Deprioritize: SEO content farms, listicles, AI-generated summaries.

2. Pick 3-5 highest-signal URLs. Use WebFetch on each:
   WebFetch(url, prompt: "Extract key claims, data points, methodology, and conclusions. Note publication date and author credentials.")

3. Cross-reference claims appearing in only one source. Flag them as unconfirmed.

## Quality rules
- Every claim needs a source citation. No unsourced assertions.
- Prioritize sources under 12 months old. Note when relying on older material.
- Distinguish facts from estimates, opinions, and projections.
- Acknowledge gaps explicitly.

## Output format

# [Topic]: Research Report

## Executive Summary
2-3 paragraph synthesis with key findings and recommendation.

## [Themed Section 1]
Findings by theme, not by source. Inline citations as `[Author/Source, Date](URL)`.

## [Themed Section 2]
...

## Actionable Takeaways
Numbered, concrete recommendations.

## Information Gaps
What could not be answered. Where sources conflict.

## Sources
Numbered list with URLs.

## Methodology
Queries run, date range of sources, domain restrictions, known blind spots.

Delivery

  • Short results (< 2000 words): the agent returns directly, relay to user.
  • Long reports: instruct the agent to save to dev/local/discovery/research-[topic-slug]-[YYYY-MM-DD].md using the Write tool (curated dir per the GC contract; never dev/local root). Tell the user where it was saved.