Skills
Multi-source deep research
for Claude Code and Cursor
anycap-deepresearch is an installable skill that gives Claude Code, Cursor, Codex, and similar coding agents a structured research workflow backed by live web data. Without this skill, an agent conducting research is limited to its training knowledge and cannot efficiently plan a search query strategy, gather from multiple live sources, deduplicate findings, and produce a structured output with source attribution. With the skill installed, the agent knows how to invoke AnyCap web search and web crawl at each stage of a research task: planning the query set, pulling full-page content from relevant URLs, filtering and cross-referencing sources, and assembling a synthesized deliverable. The workflow handles source attribution automatically so downstream documents or reports can reference original URLs. This is valuable for competitive research, technical documentation, market analysis, and any agent workflow that needs grounded, citable findings rather than model hallucination. The skill is compatible with Claude Code, Cursor, and Codex and installs in one command.
Install time
< 5 min
Supported agents
Claude Code · Cursor · Codex
Platform
macOS · Linux · Windows
Install this skill
npx -y skills add anycap-ai/anycap -s 'anycap-deepresearch' -g -y
How the workflow changes after install
BEFORE
Agent answers from training knowledge only. No live sources. No citations. Findings may be outdated or hallucinated.
AFTER
Agent plans queries, searches the live web, crawls full pages, cross-references sources, and delivers a cited report — all within one session.
- 1
Plan the query set
Agent breaks the research brief into 5–8 targeted search queries, each covering a different angle: definitions, statistics, comparisons, critiques, and recent developments. Distributing queries across angles prevents the single-query blind spot that misses counter-evidence and outlier findings.
- 2
Run web search
Agent runs `anycap web search` for each query in sequence and records titles, snippets, and URLs from live results. Running 6–8 queries ensures coverage across source types — official documentation, independent analysis, and community discussion — without over-relying on any single perspective.
- 3
Crawl top sources
Agent calls `anycap web crawl` on the top 2–3 URLs per query to pull full page content for analysis. Crawling moves beyond SERP snippets to retrieve full arguments, data tables, and contextual evidence — the kind of detail that search result previews never surface.
- 4
Cross-reference and deduplicate
Agent compares findings across all crawled pages, removes duplicate claims, and flags contradictions — preserving cases where credible sources disagree so the human reviewer can make the final judgment rather than having the agent silently resolve the conflict.
- 5
Synthesize with citations
Agent assembles a structured deliverable: findings grouped by theme, each claim attributed to a source URL. The output format adapts to the brief — Markdown for documentation, JSON for downstream processing, or prose for editorial use.
SAMPLE OUTPUT
Research complete: "context engineering for AI agents"
├── Queries run: 7
├── Sources gathered: 14 URLs
├── Full pages crawled: 9
├── Deduplicated findings: 22
└── Output: research-report.md
├── Word count: 2,840
├── Citations: 14 inline
└── Contradictions flagged: 2Capabilities used by this skill
Supported agents
Frequently asked questions
- What does this skill do that direct prompting doesn't?
- Direct prompting uses training data. This skill teaches the agent to plan a query strategy, pull live results, crawl full pages, and synthesize with citations — turning Claude into a grounded researcher rather than a text predictor.
- Can Claude Code already search the web without this skill?
- Claude Code can use MCP tools for web search, but it lacks a structured research pipeline by default. This skill adds query planning, crawl depth, deduplication, and citation formatting.
- What citation format does the output use?
- Source URLs are embedded inline with each finding. You can instruct the agent to format the final output as markdown, JSON, or plain text.
- Do I need an AnyCap account?
- Yes. Install the CLI and run `anycap login` before your first session.