Exa Deep Research Orchestrator
A skill that turns Exa search into a deep-research orchestrator: planning searches, dispatching subagents, and compiling results.
The skill requires authentication via OAuth or API key, supports anonymous access with rate limits. The skill file explicitly instructs how to handle auth failures (surface fix), no destructive operations are automatic, and data flow is querying an external search service. However, the skill does not specify least-privilege permissions, user confirmation, or rollback mechanisms. Publisher is unverified in the FollowSkills curated registry, so identity is unknown. No malicious behavior is evident, but security controls are incomplete. Deductions: missing permission scoping and confirmation mechanisms, hence 16/25.
Skill instructions are self-consistent and rely on Exa's MCP service, an external dependency with unverifiable availability. The skill provides general error handling guidance but lacks detailed failure scenarios or recovery procedures. Static review cannot execute key paths, and no third-party execution evidence (e.g., CI plus committed tests covering skill paths) is present in the skill files. Thus reliability is capped below 10. Deduction: lack of reproducible tests and detailed error handling, 8/20.
The skill is clearly tailored for deep research, lead generation, literature reviews, etc., with trigger phrases like 'research this'. It defines complexity levels and confirmation logic to avoid false triggers. However, it depends entirely on Exa's overseas service, which may be unreachable from mainland China, and provides no offline or alternative fallback, limiting its fit for FollowSkills' Chinese user base. Deduction: inadequate environment fit for mainland China, 10/15.
Documentation is well-structured with progressive disclosure, including prerequisites, steps, and troubleshooting. It includes examples and gotchas. However, there is no explicit versioning, changelog, or maintenance responsibility within the skill file itself; the skill's update path is unclear. License is MIT but publisher is not verified. Deduction: missing version/history details, 11/15.
The skill is designed to complete research tasks, but static review cannot verify actual output quality. It describes a workflow but lacks evidence that outputs are directly usable or superior to manual efforts. Static review caps effectiveness at 7 due to lack of execution evidence. Deduction: value proposition unverified, 7/15.
The repository includes CI workflows and unit tests, but these primarily cover the API handler, not the skill's execution itself. The skill file contains no verifiable third-party execution evidence. Static review cannot independently verify claims, hence 5/10.
- This skill depends entirely on Exa's overseas service, which may be unreachable from mainland China; consider alternatives.
- Publisher identity is unverified; evaluate trustworthiness independently before use.
- The skill does not specify permission scoping or confirmation mechanisms; handle sensitive data with caution.
What does this skill do, and when should you use it?
This skill transforms Exa search into an intelligent research orchestrator: it first analyzes query complexity, deciding whether to handle it directly, delegate to a single subagent, or run multiple rounds of parallel subagents. It requires the Exa MCP server, supporting OAuth, API key, or anonymous access. The skill's key value is token isolation via subagents, keeping the main context clean, and it includes detailed workflows for deduplication, source-quality evaluation, and multi-pass queries like entity chaining. It is designed for research tasks requiring comprehensive coverage, such as lead generation, literature reviews, and competitive analysis.
Reads the user query, assesses complexity (extremely simple, moderate, advanced, complex), and chooses a strategy. It instructs the orchestrator to dispatch subagents (model: haiku) using the Agent tool, each reading reference files (like references/searching.md) and running specific Exa searches. The skill defines a dispatch template including specific searches, validation criteria, and output format, and requires subagents to end with 'sources_reviewed: N'. It handles deduplication (by URL and entity merging), validates coverage with follow-up searches, and compiles results in a specific format (tables preferred, hyperlinks included), optionally writing to a file (e.g., ./exa-results/). It also guides source-quality assessment, warns against over-execution on simple queries, and asks user for clarification when ambiguous.
- Lead generation: Marketing professionals need a list of companies matching specific criteria (location, industry, size), gathered and merged from multiple sources across parallel searches.
- Literature review: Researchers need to find recent papers, preprints, and related work, filtered by date, and synthesize major themes and findings.
- Deep dive: Product managers want to understand a new topic or technology, such as WebGPU, covering specs, browser support, community adoption, and benchmarks from multiple angles.
- Competitive analysis: Startup founders want to discover competitors in a space, including funding databases, press, and niche directories, then merge and rank them.
- Entity chaining (multi-hop): Analysts need to start from a list of companies, then find people at each company, then enrich those people's public statements, building comprehensive profiles.
What are this skill's strengths and limitations?
- Structured approach to complex research tasks, with explicit differentiation between simple and complex queries.
- Token isolation from raw search results via subagents, keeping the main context window clean.
- Emphasis on validation and deduplication—ensuring search results meet user criteria, not just similarity.
- Support for multi-pass queries (e.g., entity chaining) enabling comprehensive research across sources.
- Includes source-quality assessment guidelines, weighting practitioners over commentators and defining exclusions.
- Heavily depends on the Exa MCP service, requiring network connectivity and an API key or OAuth, not a local tool.
- Instructions revolve around 'subagents' and 'Agent tool', which may be Claude Code/Anthropic-specific, limiting portability.
- Editorial involves extensive dispatch and coordination instructions, which may be overkill for simple tasks.
- No test suite or validation on non-Anthropic agents; the skill is untested on ChatGPT or Codex.
- Lacks fallback to other search engines if Exa service is unavailable or returns limited results.
How do you install this skill?
This skill is part of the exa-mcp-server repository, which bundles multiple skills. It does not provide an installation script itself; one must configure the Exa MCP server (https://mcp.exa.ai/mcp) with OAuth (recommended) or an API key for authentication. Users should place the SKILL.md file and its references/ directory into their agent's skill directory.
How do you use this skill?
Trigger the skill with phrases like 'research this', 'find everything about', or 'deep dive on'. The skill instructs the orchestrator to first calculate relevant dates, assess query complexity, and ask the user for clarification if ambiguous. For most tasks, the orchestrator will dispatch parallel subagents via the Agent tool, each using Exa's web_search_exa for specific searches. After subagents return distilled output, the orchestrator deduplicates, validates, formats, and presents results, often as tables.