ATWL Extraction & Experimentation Guide

Version: 2026.04.16

This document outlines the standard procedure for extracting, reviewing, and documenting Visual Analytics (VA) workflows from research papers using the Artifact–Transform Workflow Language (ATWL) and pre-trained LLM agents.

---

0. Preparation

• Model Selection: Select an advanced Large Language Model (LLM) with strong reasoning capabilities (e.g., Claude 3.5 Sonnet, GPT-4o, or newer).
• Parameters: Ensure the model is set to a high level of reasoning/creativity for complex structural analysis.

---

1. Phase 1: Workflow Extraction

Goal: Transform the raw text of a research paper into a structured ATWL representation.

1. Initialize Session 1: Open a new chat in your chosen LLM.
2. Load Resources: Upload the extractor.json agent file and the PDF of the research paper you wish to evaluate.
3. Prompt:

   "Please extract knowledge from the JSON log and then extract the workflow from the attached paper. Create a natural language summary and represent it in ATWL."

---

2. Phase 2: Formal Review

Goal: Audit the extracted representation for syntactic and semantic correctness.

1. Initialize Session 2: Open a new separate chat session.
2. Load Resources: Upload the reviewer.json agent file, the original research paper, and the ATWL output generated in Phase 1.
3. Prompt:

   "Please check the extracted workflow representation of the attached paper against the provided source."

---

3. Phase 3: Iterative Correction

Goal: Refine the workflow based on the reviewer's feedback.

1. Return to Session 1: Re-open the extraction session.
2. Load Input: Upload the feedback/output from the Reviewer in Session 2.
3. Prompt:

   "Here is the output from the critical assessment. Please make corrections according to the given feedback."

---

4. Phase 4: Diagram Generation

Goal: Create a publication-ready visual representation of the ATWL logic.

1. Initialize Session 3: Open a new chat session.
2. Load Resources: Upload the diagrammer.json agent file and the final corrected ATWL output from Phase 3.
3. Prompt:

   "Please generate a TikZ/LaTeX workflow diagram."

4. Refine Layout: Compile the code in a LaTeX editor (e.g., Overleaf). If the layout is unsatisfactory, capture a PDF or screenshot and upload it back to Session 3.
5. Refinement Prompt:

   "Improve the layout and arrow placement. You are allowed to simplify and/or represent at a higher level of abstraction for better readability."

---

5. Finalizing the Experiment

• Success: If you are satisfied with the results, save your sessions for the Extractor, Reviewer, and Diagrammer. These sessions have "learned" from your specific experiment and can be reused for similar papers.
• Reset: If the results were not satisfactory, discard the sessions and start fresh with the original JSON agent files next time.

Enjoy your experiments!