Auto_Analysis — System Overview Project Description

Auto_Analysis is an automated data analysis workflow system.

A user provides:

a natural language goal (what they want to learn), and one or more datasets (CSV, Excel, JSON, etc.)

The system then autonomously:

explores the data selects appropriate analytical methods executes code iteratively refines its approach produces a final report

A large language model (Claude) is responsible for reasoning, planning, and choosing analytical strategies, while the surrounding system acts as a harness that provides tools, enforces structure, and records all steps.

Core Workflow User Goal + Data ↓ Planner / Coordinator ↓ Claude Reasoning Step ↓ Decide Next Action ↓ ┌──────────────────────────────┐ │ Can we proceed with action? │ └──────────────────────────────┘ ↓ yes ↓ no Tool Execution Ask User (Clarify) ↓ ↓ run_python / SQL Receive User Input ↓ ↓ Observe Results ←───────────────┘ ↓ Update State / Memory ↓ Re-plan Next Step ↓ (Loop until analysis is complete) ↓ Generate Final Report Agent Responsibilities (Claude)

Claude is responsible for:

Interpreting the user’s goal Inspecting dataset structure Forming hypotheses about the data Selecting appropriate analytical methods (EDA, regression, classification, clustering, etc.) Writing analysis code (Python / SQL) Interpreting execution results Deciding when clarification is required Producing a final human-readable report System Responsibilities (Harness)

The surrounding system is responsible for:

Providing a controlled execution environment (run_python, run_sql) Managing file/data access Enforcing step-by-step workflow structure Capturing execution outputs and errors Maintaining session state and analysis history Handling user clarification prompts (ask_user) Logging all actions for reproducibility Key Design Principle

Claude performs reasoning. The system enforces execution, structure, and reliability.

This separation ensures:

consistent analytical behavior real execution of code (not hallucinated results) iterative improvement through feedback loops traceable, reproducible workflows

If you want, I can next turn this into a 1-page “README.md style spec” or a diagram-ready architecture block for your repo.