2 Getting Started
When adopting agentic workflows, success is determined mostly by whether you simply start using the tools or not.
The intricate mechanics and underlying details of how these tools work can come later. You do not need to master every concept upfront before getting value from them—especially given the fact that you can ask the AI to set itself up, configure your environment, and explain how everything works in real time.
The intuition, knowledge, and practical skills required to steer agentic AI will grow organically through daily practice and active experimentation. It is far better to start using the tools as soon as possible on real research tasks rather than spending weeks memorizing abstract theories, frameworks, and terminology.
To begin on Day 1, you fundamentally just need three things:
- Install an agentic AI managing software (the “harness”)
- Create a project folder (or clone this repository)
- Start a conversation
2.1 Choosing and Installing Your Harness
You can choose any of the leading agentic coding harnesses. There is no need to purchase a new tool if you already have access to one—simply pick whichever software harness corresponds to the subscription you already own, or whichever platform is licensed and provided by your university, school, or research institution.
While this guide uses Google Antigravity as its primary illustration and certain advanced features are tailored to it, the foundational workflow habits, prompt patterns, and plan-first principles translate well across major platforms:
- Google Antigravity: Google’s agentic development environment and desktop application powered by Gemini frontier models.
- Claude Desktop: Anthropic’s official desktop application with local project integration.
- ChatGPT Desktop / Codex: OpenAI’s official desktop application and workspace environment.
- Use What You Already Have: If you or your university already have a subscription (such as Google Workspace with Gemini / Gemini Advanced, ChatGPT Plus/Team/Edu, or Claude Pro/Team), start with that software harness.
- Free Tiers: If you don’t have an active subscription yet, both Google Antigravity and OpenAI provide accessible free tiers to help you experiment with agentic workflows immediately at zero cost.
- Recommended Approach (Subscriptions): For researchers setting up a personal account, we strongly recommend a standard monthly subscription rather than managing complex API keys and pay-per-token billing. Subscriptions provide predictable monthly costs, access to top-tier frontier reasoning models, higher usage allowances, and a seamless interface.
The tools listed above all have a built-in sandbox environment. By default, the agent’s actions and file access are strictly scoped to your active project directory, ensuring that your other personal files and system directories on your computer will not be touched without your explicit permission.
2.2 Creating a New Project in Your Harness
Once you have installed your preferred desktop application, follow these simple steps to set up your project workspace:
- Launch Google Antigravity on your computer.
- Open Your Project Directory: Click “Open Folder” on the startup screen (or choose
File > Open Folder...from the top menu). - Select Your Folder: Choose the folder where you cloned this template repository (or select any research directory on your machine).
- Automatic Workspace Setup: Antigravity immediately opens the workspace, scopes its secure sandbox to that folder, and detects existing configuration files (like
GEMINI.mdand.agents/). - Start Conversing: The agent chat panel is ready immediately—you can start conversing with the AI right away!
- Launch Claude Desktop on your computer.
- Navigate to Projects: Click on “Projects” in the left sidebar (available with Pro/Team/Edu subscriptions).
- Create a Project: Click “Create Project” (or the
+icon) and give it a name (e.g.,CEO-Turnover-EmpiricsorAccounting-Theory-Paper). - Attach Files & Context: Add your paper drafts, data files, or project guidelines to the project context. If using local filesystem connectors, link your project folder directly.
- Start Conversing: Open a chat within that project to begin collaborating and chatting immediately.
- Launch ChatGPT Desktop on your computer.
- Open Projects / Workspaces: In the sidebar, select “Projects” (or use the Workspace / Canvas feature in Plus/Team/Edu plans).
- Create New Project: Click “New Project”, name your research topic, and set your project-specific custom instructions.
- Sync Project Files: Attach your working scripts, notes, or LaTeX documents to the project workspace.
- Start Conversing: Open a chat inside the project to immediately begin the conversation.
2.2.1 You Can Start Conversing Right Away!
Once your project workspace is open, you do not need to do any complex setup before getting started. You can immediately begin chatting with the AI—ask it questions, discuss your research ideas, or direct it to help draft your materials.
2.2.2 Why Use a Developed Setup Instead of an Empty Folder?
While you can always start with an empty folder, using a developed setup offers key advantages:
- Eliminates Repetitive Prompting: You don’t have to repeatedly re-teach the AI your formatting preferences, notation rules, or coding standards across sessions.
- Built-in Quality & Error Prevention: The AI is pre-calibrated to follow rigorous review habits, catching typos, formatting flaws, and subtle errors automatically.
- Cross-Session Memory: It preserves project decisions, notation preferences, and learned corrections so you never have to start from scratch each day.
- Instant Productivity: You can skip hours of manual setup and immediately focus on the substantive research questions and creative ideas.
2.3 Recommended Companion Tools
You likely already have your preferred empirical software (R, Stata, or Python) installed—your AI agent will connect directly to your existing environment. To streamline your collaboration with the agent, here are a few recommended companion tools:
- Integrated Development Environment (IDE): VS Code
- Why use an IDE for your research projects? An IDE provides a clear visual picture of your entire project folder structure and gives you instant access to all scripts, datasets, and AI-generated files in one centralized window.
- Integrated Harness Add-ons: You can install official extensions for your chosen AI harness directly from the VS Code Marketplace, letting you converse with the agent alongside side-by-side code diffs.
- Built-in File Previews: By installing simple add-ons (such as a PDF viewer extension), you can preview compiled Beamer decks and papers directly inside the IDE without juggling external reader apps.
- Version Control: Git and optionally GitHub Desktop for project snapshots, safety checkpoints, and change history.
- Voice-to-Text Dictation: Pipit Voice (macOS) for local, zero-delay speech-to-text. Dictating instructions at natural speaking speed is significantly faster and more responsive than the built-in voice tools in most AI harnesses.
- Document & Presentation Compilers: MacTeX / MiKTeX for LaTeX/Beamer and Quarto CLI for web slides and reports.
You don’t need to configure your environment alone. You can copy and paste these two prompts into your AI agent:
Prompt 1: Setup & Guide for Desktop / GUI Applications (Non-CLI)
Please check my operating system and verify if the following desktop software is installed: VS Code, GitHub Desktop, MacTeX / MiKTeX, and Pipit Voice. For any missing software, provide the official download link and guide me step-by-step through the installation and initial configuration.Prompt 2: Automated Installation for Command-Line Tools & Packages (CLI)
Please run terminal checks to see if Git, Quarto CLI, R, and Python are available on my machine. If any command-line tools or package managers (such as Homebrew on macOS or winget on Windows) are available, propose and execute the commands to install the missing tools and verify their versions for me.2.4 Adopting an Established Academic Setup
You do not need to build rules, prompt libraries, and configuration files from scratch. The academic community has developed ready-to-use setups tailored to specific harnesses:
- For Claude Code / Claude Desktop Users:
- The vast majority of community academic setups in economics and finance have been built for Claude Code. Leading examples include Barrios88/barrios-skills by John Barrios (Washington University in St. Louis) for empirical economics, WRDS access, SEC EDGAR scraping, and slide workflows, and hugosantanna/clo-author by Hugo Sant’Anna for paper-centric multi-agent writing and referee reviews.
- For Google Antigravity Users:
- GangPan94/AntigravitySetup — The companion setup repository for this guide, tailored specifically to bring these academic workflows,
GEMINI.md, 70+ academic skills, multi-agent adversarial reviews, and memory management into the Google Antigravity environment.
- GangPan94/AntigravitySetup — The companion setup repository for this guide, tailored specifically to bring these academic workflows,
- For OpenAI Codex / ChatGPT Users:
- Because almost all open academic setups are currently created for Claude Code, OpenAI provides an official, curated migration skill called
migrate-to-codex(part of the officialopenai/skillsrepository). - This tool automates the transition of instructions (
CLAUDE.md), skills, and configuration files from a Claude Code repository (such as Barrios88/barrios-skills) directly into native Codex format (AGENTS.mdand.codex/configurations) with self-healing dry runs and validation. Alternatively, you can use the prompt below to let your Codex agent perform the conversion directly.
- Because almost all open academic setups are currently created for Claude Code, OpenAI provides an official, curated migration skill called
2.4.1 Forking the Setup: Your Own Personal Copy
Before importing a setup, you will need a free GitHub account. We strongly recommend that you fork the original repository to your own account first, and then provide the link to your fork.
Think of a fork as making your own personal photocopy or working duplicate of someone else’s master binder.
When you click the “Fork” button (located near the top-right corner of any GitHub repository page), GitHub creates an exact, independent copy of that entire project under your personal account (https://github.com/your-username/repository-name).
Why fork instead of using the original link directly?
- Freedom to Customize: You can modify rules, add personal preferences, and change prompts without worrying about altering the original author’s repository.
- Easy Upstream Updates: Whenever the original author releases new skills, bug fixes, or improved prompts, GitHub allows you to click “Sync fork” to pull in those updates with a single click.
- Cloud Backup: Your customized academic setup is safely preserved in your own GitHub account across computers and research sessions.
2.4.2 The “Zero-Effort” Setup Prompt
Once you have forked the repository, simply open your project folder in your chosen harness, paste the prompt below into chat, and provide the link to your forked repository (https://github.com/your-username/your-repo-name). The AI will configure and convert everything for you automatically:
Please configure this project workspace by importing the academic setup and skills from my repository: **[INSERT YOUR FORKED REPO LINK, e.g., https://github.com/your-username/AntigravitySetup]**.
Inspect the repository structure, copy over the skills (.agents/skills/ or .claude/skills/), preambles, and memory scaffolding into my active project folder, and convert the project constitution (such as CLAUDE.md or GEMINI.md) into the appropriate format for my current harness (e.g., AGENTS.md for Codex / ChatGPT, CLAUDE.md for Claude Code, or GEMINI.md for Antigravity). Do not overwrite or modify any existing research data files.To browse more field-specific setups, domain skill libraries (e.g., theoretical economics, presentation design, automated replication), or empirical data tools, see Appendix: Academic AI Setups & Ecosystem.
2.5 The Starter Prompt
When starting a project or initializing a fresh workspace, copy and paste the prompt below into your agent chat. Fill in the bolded placeholders:
I am starting to work on **[PROJECT NAME]** in this repo.
**[Describe your project in 2–3 sentences — what you're building, who it's for, what tools you use (e.g., LaTeX/Beamer, R, Quarto, Stata).]**
I want our collaboration to follow the Agentic Workflow paradigm: plan-first, multi-agent verification, adversarial QA loops, and mandatory compilation checks before any task is considered complete. I don't want to repeat myself — our workflow should be smart about remembering decisions and learning from corrections.
The configuration files are in `.agents/skills/`, `.agents/agents/`, `.agents/rules/`, and `.agents/references/`. Please read them, understand the workflow, and **update all configuration to fit my project** — fill in placeholders in `GEMINI.md` (or `CLAUDE.md` / `AGENTS.md`), adjust rules if needed, and propose any customizations specific to my use case.
Use the plan-first workflow for all non-trivial tasks. Once I approve a plan, act as the Orchestrator: coordinate everything autonomously using subagents where appropriate, only pausing when there is genuine ambiguity or a decision to make.
Enter plan mode and start by adapting the workflow configuration for this project.2.5.1 Why Are We Doing This?
While the prompt’s instructions are straightforward, one critical and non-obvious benefit is preventing tool-routing confusion and instruction bloat:
- Tailoring Available Skills: Comprehensive academic setups often come with dozens of domain skills (e.g., WRDS financial access, Stata replication scripts, theorem proving, Beamer slide builders). Stating your project’s scope upfront allows the AI to focus on the skills relevant to your specific methodology and ignore unneeded tools.
- Accurate Tool Routing: A curated, project-aligned configuration prevents the model from invoking the wrong tool or getting distracted by unrelated capabilities during complex research tasks.
- Staying Within the Instruction Budget: Frontier language models reliably follow around 100–150 custom instructions at a time. Replacing generic template placeholders and tailoring the active rules ensures your project’s constitution (
GEMINI.md/CLAUDE.md/AGENTS.md) remains lean and strictly followed.
(We will revisit how skills, path-scoped rules, and context windows work in detail in Chapter 5).
2.6 Requirements Specification & /grill-me
For complex, creative, or ambiguous tasks (such as framing an empirical identification strategy, scoping a theoretical model, or structuring a new course syllabus), jumping straight into execution often leads to costly rework. The spec-then-plan pattern catches ambiguity before any files are created or modified.
2.6.1 The /grill-me Interactive Interview
When you have a rough idea that needs crystallization, you can invoke the /grill-me command (or prompt your agent: “Grill me on this research idea before creating a plan”):
- One Question at a Time: The agent conducts a structured, multi-turn interview—asking probing questions about your research question, competitor differentiation, identification strategy, or methodology scope.
- Surfacing Hidden Constraints: It pushes you to resolve critical design decisions and trade-offs upfront rather than encountering roadblocks mid-implementation.
2.6.2 The Specification Document
Once the interview or clarifying questions are complete, the agent synthesizes your answers into a structured specification document saved to quality_reports/specs/ containing:
- MUST have: Non-negotiable requirements (e.g., specific estimators, sample filters, baseline controls)
- SHOULD have: Preferred features and secondary robustness tests
- MAY have: Optional extensions or exploratory analyses
- Clarity status:
CLEAR,ASSUMED, orBLOCKEDfor each component
You review and approve this specification before planning begins. This spec-then-plan sequence drastically reduces mid-project pivots and ensures you and your agent are perfectly aligned.
If you have already installed your harness, opened your project folder, and directed your agent to execute your very first task, congratulations—you are already 80% of the way toward mastering this workflow!
The remaining chapters will dive deeper into the architectural mechanics, multi-agent adversarial loops, and specialized research patterns, but the foundational habit of active, plan-first collaboration is already well underway.