ResearchBot: evidence-based research for Claude Code

ResearchBot is a Claude Code skill and research agent, built at AR Labs, that turns a question into findings a reader can check. It is for research that will inform a real decision: it evaluates each source before trusting it, looks for independent confirmation of major claims, gives contradicting sources equal space, and says plainly where the evidence runs out. Each finding carries a confidence grade, and the report lists the searches that failed to find what was needed. We use it when the credibility of a source matters more than the speed of an answer; for quick fact-finding we use multi-agent research instead.

How it works

A request names a depth and a topic, for example /researchbot D2 multi-agent orchestration frameworks; without a depth it runs at D2. The command is a thin dispatcher: it hands the work to a dedicated research agent (a subagent, which in Claude Code is a separate agent session with its own instructions and tool list), one per pass at the deeper levels. That agent is a leaf worker: its tool list contains web search, page fetching, file reading and writing, but not the tool that starts other agents, so it cannot delegate further. The limit is set in the tool list rather than written as a request in the prompt.

Research runs in passes that deepen progressively:

• D1, focused: Pass 1 only, up to 5 searches and 15 sources, one short document.
• D2, standard: Passes 1 and 2, up to 10 searches and 30 sources, one to three documents. This is the default.
• D3, comprehensive: all three passes, up to 15 searches and 50 sources, several documents with an index. Pass 3 runs only if Pass 2 found critical gaps, and only after the person who asked approves the specific areas it will cover.

Pass 1 maps the landscape: the main themes, the authoritative sources, and the contrarian views. Pass 2 follows the most promising pathways in depth. Synthesis groups coded claims into themes, then moves from description to analysis.

The judgement it encodes

Score the source before using it. Every source gets a CRAAP score. The CRAAP test is a source-evaluation checklist devised by Sarah Blakeslee, a librarian at California State University, Chico: Currency, Relevance, Authority, Accuracy and Purpose. ResearchBot rates each 0 to 5 and weights them, with authority counting most (40 per cent) and purpose least (5 per cent). A source scoring 4.0 or more may support a major claim; 3.0 to 3.9 needs verification; below 3.0 is reported as a fringe view.

Read laterally. Following the practice of professional fact-checkers, the agent spends little time on the source itself and more on what independent assessments say about it.

Triangulate major claims. A major claim needs three independent sources, meaning different authors and institutions that do not cite each other. When sources disagree, both views are documented with equal prominence, the possible reasons are investigated, and the finding is marked disputed.

Grade confidence on every finding. High, moderate, low, or insufficient evidence, matched to the strength of what was found.

Document the gaps. Each evidence gap lists the queries that were tried and what would fill it, so a reader can tell "not found" from "not looked for".

Know when to stop. Searching ends when about four in five sources agree, or when five consecutive sources add less than a tenth that is new. If a run takes more than twice its expected time, the agent stops and reports partial results rather than padding them.

What exists today

ResearchBot has produced more than a dozen reports in our research library, several of which are summarised in this site's research section. It is not released as a separate package. The CRAAP score is the agent's own structured judgement, not a calibrated probability, and the confidence grades describe the evidence found rather than guarantee accuracy. The arithmetic is not always the weighted formula: in the example below, the 4.2 is the plain average of the five ratings, where the weighting would give 4.05.

Example

One of our D2 reports asked whether one language model reviewing another improves quality and readability, and whether the habit of writing as if the reader shares the model's context has an established name. Its findings are summarised in Using one LLM to review another. Abridged, with one citation corrected (the paper is by Shaikh et al.):

# LLM-Checks-LLM: Quality and Readability
Research depth: D2 | Searches: 10/10 | Sources: ~32
 
## Q1: The Unshared-Context Writing Pattern
### Is It Formally Named?
No single established term maps precisely to this pattern in either
NLP or agentic-AI literature as of 2026-08.
 
| Candidate term                          | Status                         |
|-----------------------------------------|--------------------------------|
| Common ground / grounding failure (Clark) | Nearest established concept  |
| Curse of knowledge (LLMs)               | Studied in LLM judges, not writers |
| Context leakage / context bleed         | Security literature; not this pattern |
 
1. "Grounding Gaps in Language Model Generations", Shaikh et al. (NAACL 2024)
CRAAP: Currency 3, Relevance 5, Authority 4 (peer-reviewed, ACL),
Accuracy 4, Purpose 5 -> CRAAP 4.2
 
Confidence: Moderate. The phenomenon is documented at the mechanism
level. It is not yet formally named in the NLP or agentic-AI
literature.
 
Evidence gap: Searched with 6 framings ("unshared context",
"context leakage", "audience model", "reader model", "presupposition
failure", "egocentric bias") and found no paper with this exact
failure as its primary topic. The gap is real.