- Home
- Skills
- Research & Analysis
- STORM Research
More screenshots
Works with the AI Agents you already use
STORM Research
Source-grounded research with cited HTML reports, traceable evidence, and a prompt-native Co-STORM preview.
Free
STORM Research
Example session with this skill installed
Use the storm skill to research RAG evaluation frameworks. Produce the standard HTML artifact bundle and focus on how an engineering team should choose among RAGAs, ARES, and RAGChecker.
- Read your context and instructions
- Compiled the storm research
Repository example, generated 2026-07-10. The complete polished article is reproduced below in Markdown. This is an existing source-grounded output snapshot, not an Agensi execution test or benchmark. View the original HTML and four-file progression.
Polished artifact · evidence checked
Choose RAG evaluation by the failure you need to see
RAGAs, ARES, and RAGChecker overlap on relevance and faithfulness, but they serve different evaluation moments: fast feedback, calibrated comparison, and component diagnosis.
Generated 2026-07-10 · Four primary papers · No experiment reproduction
The system boundary creates the evaluation boundary
The original RAG architecture combines a parametric generator with retrieved non-parametric memory and presents provenance and updateable knowledge as core motivations.[1] That modularity means a final answer score cannot explain the whole system. Teams need evidence about retrieval quality, the generator's use of context, and the quality of the answer seen by the user.
Three frameworks, three useful emphases
RAGAs targets reference-free multidimensional evaluation. Its paper distinguishes relevant and focused retrieval, faithful use of retrieved passages, and generation quality while avoiding a requirement for human ground truth on every evaluated case.[2]
ARES uses synthetic data to train lightweight judges for context relevance, answer faithfulness, and answer relevance. It then uses a small human-annotated set with prediction-powered inference to mitigate judge error. The paper evaluates the approach on eight knowledge-intensive tasks and studies domain shifts.[3]
RAGChecker moves from scoring toward diagnosis. It defines fine-grained retrieval and generation metrics, applies them to eight RAG systems, and reports stronger correlation with human judgments than the comparison metrics in its meta-evaluation.[4]
| Primary need | Method-aligned starting point | What still needs validation |
|---|---|---|
| Fast regression feedback across several RAG dimensions | RAGAs | Judge sensitivity and fit to the team's domain |
| Calibrated comparison with a small labeled evaluation set | ARES | Synthetic training assumptions and annotation design |
| Localization of retrieval and generation failure patterns | RAGChecker | Metric interpretation and transfer to the deployed architecture |
Practical synthesis. Use automated multidimensional checks for routine regressions, escalate surprising changes to fine-grained diagnosis, and retain human review for consequential or domain-specific judgments. This is an engineering recommendation inferred from the frameworks' published designs, not a result reported by any one paper.
References
- Lewis et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020.
- Es et al. RAGAs: Automated Evaluation of Retrieval Augmented Generation. EACL 2024.
- Saad-Falcon et al. ARES: An Automated Evaluation Framework for Retrieval-Augmented Generation Systems. NAACL 2024.
- Ru et al. RAGChecker: A Fine-grained Framework for Diagnosing Retrieval-Augmented Generation. NeurIPS 2024.
Verification notes
- Citation order follows first appearance; every citation maps to one primary source.
- Framework descriptions were checked against the linked NeurIPS or ACL Anthology abstracts on 2026-07-10.
- The article does not claim independent reproduction, cost comparison, latency comparison, or universal framework superiority.
- Selection guidance is explicitly marked as an inference from method design.
- Later software changes may not be reflected in the cited 2024 papers.
Connects securely to your tools. The creator never sees your data.
About this skill
STORM Research turns a research question into a structured, source-grounded report using multiple perspectives, simulated interviews, evidence-refined outlines, inline citations, and verification notes.
What it produces
Classic STORM produces four HTML files: a topic-only outline, an evidence-refined outline, a cited draft, and a polished article with references. Bundled Python scripts manage staged output, run state, citation mappings, artifact checks, and publication hashes. These checks support review; they do not independently prove that every claim is true.
Use it for literature reviews, technical background reports, source comparisons, and research over a user-provided local corpus. A chat-only brief is available with reduced validation.
Interactive and experimental modes
- Co-STORM is a prompt-native preview: simulated participants, a cited mind map, and free-form user steering. This package does not bundle the upstream CoStormRunner, DSPy modules, or independently running expert agents.
- Retrieval routing is experimental: host-ranked evidence or deterministic local lexical retrieval.
- Official Classic STORM output import is experimental: it imports a compatible existing run through guarded phases. It does not install or execute knowledge-storm. A separately captured polished-reference map is required.
Requirements and compatibility
The full file-producing workflow requires an agent that can read SKILL.md and its companion files, execute Python scripts, and read and write files in an authorized local workspace. Online research additionally needs the host agent's search or retrieval tools. The bundled runtime uses the Python standard library without third-party Python dependencies.
Agensi runtime compatibility has not been verified. Listing or ZIP parsing on Agensi does not establish successful execution of the complete workflow. Tool availability and automatic triggering vary by host.
Package and license
This free v0.6.1 package contains SKILL.md, 5 Python scripts, 10 reference and contract files, agent metadata, and the original MIT LICENSE. Copyright (c) 2026 lizhouai. Preserve the copyright and permission notice when redistributing.
Based on the research pattern from Stanford STORM. Source, usage instructions, examples, and release history: lizhouai/storm-research-skill.
Agensi runtime compatibility has not been verified. The complete guarded workflow needs Python execution and file read/write access. Guarded Classic output supports HTML only; chat-only fallback cannot mechanically enforce the same transitions, recovery, hashes, or artifact gates. Co-STORM is a prompt-native preview with simulated participants, not the upstream CoStormRunner. Retrieval routing and official Classic output import are experimental. The importer supports compatible existing stable knowledge-storm >=1.1.1,<1.2 Classic output and requires a separate polished-reference map. Citation checks validate structure and mappings; source support still needs agent or human review.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
3 uses
Times used to date
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 1 month ago
- Free with an Agensi account
Permissions
File Scopes
Terminal / Shell is used to execute the bundled Python scripts. Read Files is needed for skill references, user-authorized source material, and existing runs. Write Files is needed for run state, staging, audits, and the four HTML outputs under .results/<topic-slug>/. Browser / Network Access applies only when the host agent retrieves online sources; local-corpus work remains within the supplied boundary. The bundled retrieval and import scripts do not contact the network. No connected account or environment-variable access is required by the core package. Installing or executing a separate upstream runner requires explicit user authorization.