Deep Research
PassAudited by ClawScan on May 16, 2026.
Overview
This is an instruction-only research workflow that uses delegated subagents, web research, and optional external model checks, with no artifact-backed evidence of hidden code, exfiltration, purchases, or destructive behavior.
This skill appears safe to install as an instruction-only research assistant. Before using it, be aware that it delegates work across multiple research agents, may use web/API sources, may process supplied corpus notes, and can optionally involve another model for critique if configured. Avoid providing confidential research materials unless you are comfortable with those flows.
Findings (3)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
Your research topic and any supplied materials may be processed by multiple delegated research agents using file and web tools.
The skill explicitly delegates research tasks to subagents with file and web access. This is central to the stated research purpose, but it means subagents may access provided files and web sources during research.
The main Hermes agent is the ORCHESTRATOR... Each research phase spawns subagents via `delegate_task` ... toolsets=["file", "web"]
Use it for research tasks where web lookup and file reading are appropriate, and only provide files or corpus data you intend the research pipeline to use.
If cross-model review is enabled, drafts or research materials could be shared with another model service for critique.
The Devil’s Advocate agent documents an optional cross-model review path that can send reviewed material to another model or provider. This is disclosed and optional, but the destination/provider is not specified in the quoted artifact.
When `ARS_CROSS_MODEL` is set, after completing each checkpoint report, send the reviewed material ... to the cross-model for an independent critique.
Only enable cross-model review if you understand and trust the external model/provider, especially for confidential, unpublished, or proprietary research.
User-supplied corpus entries and notes may shape search coverage, inclusion decisions, and bibliography output.
The bibliography workflow can use a user-supplied literature corpus and notes as research context. The artifact includes safeguards such as read-only corpus handling, but stale or incorrect corpus entries can still influence results.
When the input Material Passport carries a non-empty `literature_corpus[]`, this agent enters the corpus-first, search-fills-gap flow... Read ... optional fields present (`venue`, `doi`, `tags`, `abstract`, `user_notes`).
Review any provided literature corpus for relevance and accuracy before using it, and check the generated inclusion/exclusion decisions.