░█████╗░██████╗░██╗░█████╗░
██╔══██╗██╔══██╗██║██╔══██╗
███████║██████╔╝██║███████║
██╔══██║██╔══██╗██║██╔══██║
██║░░██║██║░░██║██║██║░░██║
╚═╝░░╚═╝╚═╝░░╚═╝╚═╝╚═╝░░╚═╝
An AI-powered research automation system built entirely inside Claude.ai — no backend, no IDE, no deployment. Just prompts.
ARIA (Autonomous Research Intelligence Agent) is a fully autonomous AI research assistant that discovers, analyzes, and stores academic AI research papers — automatically, every day.
It was built entirely through a single Claude.ai conversation using structured prompt engineering, the Notion MCP connector, and Claude's built-in web search tool. No separate backend. No API integrations. No deployment pipeline.
Every day at 10:00 AM IST, ARIA:
- Searches arXiv, Semantic Scholar, and Google Scholar for high-impact AI papers
- Analyzes each paper using the ReAct reasoning framework
- Checks for duplicates before storing
- Creates a Notion database entry with 14 structured properties
- Writes a full analysis page inside each entry — Summary, Key Findings, Technical Contribution, Future Vision, and more
- Rates each paper on an Importance Score (1–10)
| Feature | Description |
|---|---|
| 🔬 Daily Research Cycle | Discovers 3 high-impact AI papers every day across arXiv, Semantic Scholar, Google Scholar |
| 🗄 Notion Knowledge Base | Creates and populates the AI Research Papers database with 14 properties and 5 smart views |
| 📄 Rich Analysis Pages | Each paper gets a full structured page — Summary, Key Findings, Technical Contribution, Future Vision, Related Papers, Personal Notes |
| 🔍 Deduplication | Checks if a paper already exists before creating an entry — no duplicate rows |
| ⏰ IST Scheduler | Auto-runs at 10:00 AM IST daily with a live countdown timer and run history |
| ⚙️ Agent Execution Trace | Live step-by-step log of ARIA's reasoning — Plan → Search → Analyze → Store |
| 🛡️ Self-Critique Mode | ARIA verifies its own outputs before finalizing — papers real? citations accurate? |
| 🔖 5 Smart Notion Views | Latest Research · Most Influential · AI Agents · Deep Dive Queue · High Impact Papers |
┌─────────────────────────────────────────────────────────┐
│ ARIA Stack │
├──────────────────┬──────────────────────────────────────┤
│ Frontend │ React (JSX Artifact in Claude.ai) │
│ Agent Brain │ Claude claude-sonnet-4-20250514 │
│ Web Research │ web_search_20250305 tool │
│ Knowledge Base │ Notion MCP Connector │
│ Scheduler │ JavaScript setInterval (IST) │
│ Memory │ Conversation history array │
│ Deduplication │ Title-match in storage prompt │
└──────────────────┴──────────────────────────────────────┘
This entire system was built through conversation — no code editor, no terminal, no deployment.
01 PROMPT ENGINEERING
Designed ARIA's system prompt using ReAct framework,
AutoGPT-style 8-step loop, and strict output format templates.
02 WEB SEARCH TOOL
Connected Claude's built-in web_search_20250305 tool for
real-time discovery from academic databases.
03 NOTION MCP CONNECTOR
Enabled the Notion MCP connector in Claude settings.
ARIA can create databases, write pages, and manage entries.
04 DATABASE ARCHITECTURE
Designed a 14-property Notion schema. Each paper = one row
+ a rich analysis page inside that row.
05 AGENT LOOP
Implemented AutoGPT-style 8-step loop with deduplication
built into the storage step.
06 SCHEDULER
Built a browser-based IST scheduler. At 10:00 AM IST,
ARIA auto-triggers and populates the Notion database.
ARIA's behavior is entirely defined by its system prompt — structured like a program:
┌─────────────────────────────────────────────┐
│ ARIA System Prompt │
├─────────────────────────────────────────────┤
│ Identity Layer → Name, role, persona │
│ ReAct Framework → Reason→Act→Observe │
│ AutoGPT Loop → 8-step task cycle │
│ Output Template → Exact report format │
│ Notion Rules → DB + page structure │
│ Page Template → 8-section analysis │
│ Constraints → No hallucinations │
│ Self-Critique → Verify own outputs │
└─────────────────────────────────────────────┘
Key insight: Treating the prompt like a program — with specific rules, templates, and constraints — produces reliable, autonomous behavior.
The AI Research Papers database has 14 properties:
| Property | Type | Purpose |
|---|---|---|
| Title | title |
Primary key — paper name |
| Authors | rich_text |
Author names |
| Year | number |
Publication year |
| Publication | select |
arXiv / NeurIPS / ICML / ACL / ICLR / CVPR |
| Paper Link | url |
Direct link to paper |
| Citation Count | number |
Influence metric |
| Research Category | multi_select |
LLM Architecture, AI Agents, RAG, Safety… |
| Key Idea | rich_text |
One-sentence core insight |
| Significance | rich_text |
Why it matters now |
| Future Impact | rich_text |
Long-term implications |
| Date Added | date |
Auto-set by ARIA |
| Status | select |
New / Reviewed / Deep Dive |
| Importance Score | number |
ARIA's rating 1–10 |
| View | Logic |
|---|---|
| 📊 Latest Research | Sort by Date Added ↓ |
| 🔥 Most Influential | Sort by Citation Count ↓ |
| 🤖 AI Agents Research | Filter: Category contains "AI Agents" |
| 📚 Deep Dive Queue | Filter: Status = "New" |
| 🚀 High Impact Papers | Filter: Importance Score ≥ 8 |
Every database row opens as a full Notion page with this structure:
# Research Paper Analysis
## 📄 Paper Summary
High-level explanation of the research.
## 🔍 Key Findings
• Finding 1
• Finding 2
• Finding 3
## 🔬 Technical Contribution
What new method or innovation this paper introduces.
## ⚡ Why This Research Matters
Importance in the current AI ecosystem.
## 🔭 Future Vision
Possible long-term impact on AI development.
## 💡 Key Takeaways
• Takeaway 1
• Takeaway 2
## 📚 Related Papers
Papers worth exploring next.
## ✏️ Personal Research Notes
← Your own thoughts go hereARIA tags every paper with one or more of these multi-select categories:
LLM Architecture AI Agents
Retrieval Augmented Generation AI Infrastructure
Model Optimization Alignment & Safety
Multimodal AI Robotics AI
Training Methods Evaluation Methods
ARIA uses the ReAct (Reasoning + Acting) framework for every research task:
┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐
│ REASON │ ──► │ ACT │ ──► │ OBSERVE │ ──► │ REFLECT │
│ │ │ │ │ │ │ │
│ Think │ │ Search │ │ Review │ │ Refine │
│ about │ │ analyze │ │ results │ │ & store │
│ task │ │ retrieve│ │ │ │ │
└─────────┘ └─────────┘ └─────────┘ └─────────┘
Combined with the AutoGPT-style task loop:
Plan → Discover → Filter → Analyze → Synthesize → Reflect → Store → Learn
Before finalizing any report, ARIA verifies its own outputs:
SELF-CRITIQUE CHECK
├── Are all papers real? [YES / UNCERTAIN]
├── Are citations accurate? [YES / UNCERTAIN]
├── Are summaries faithful? [YES / UNCERTAIN]
└── Are insights logically derived? [YES / UNCERTAIN]
If uncertainty exists, ARIA explicitly flags it in the output rather than presenting guesses as facts.
ARIA is available as a free prototype inside Claude.ai — no API key needed. The artifact calls the Anthropic API through Claude's built-in access.
To use it:
- Open claude.ai
- Load the ARIA artifact from this repository
- (Optional) Enable the Notion MCP connector in Claude settings for database storage
- Click "🔬 Run Daily Research Cycle" and watch ARIA work
ARIA/
├── ARIA.jsx # Main React artifact (free prototype)
├── ARIA_Standalone.html # Standalone HTML version (requires API key)
├── ARIA_Portfolio_Case_Study.docx # Full build documentation
├── ARIA.mp4 # Prototype video
└── README.md # This file
Building ARIA taught me that prompt engineering is software engineering — when done with the same rigor:
- Specificity beats vagueness — every rule in ARIA's prompt is unambiguous
- Structured output templates make responses machine-parseable, not just human-readable
- MCP connectors are superpowers — they turn a chatbot into a real automation system
- Iterative refinement works — the final system emerged through 6+ rounds of conversation
- The database-as-knowledge-base pattern — each row opening as a full analysis page is far more valuable than a flat log
A system that would normally require a backend server, a database, multiple API integrations, and weeks of engineering — built entirely in conversation.
Structured Prompts + MCP Connectors + Built-in Tools + Iterative Refinement
=
A deployable autonomous research agent
- Claude AI — Anthropic's AI assistant
- Notion — Knowledge base and database
- Notion MCP — Model Context Protocol connector
- ReAct Paper — Reasoning + Acting framework
- arXiv — Open-access research repository
Built with Claude AI · No backend · No IDE · Just prompts
"The best architecture is the one you can explain to a language model."