AI Research Setup

A step-by-step guide to setting up AI tools for research, on Windows and macOS.

Five steps, in order: install → configure → workflow → software → backup. Each step keeps the essentials in this README and links to a sub-page in docs/ for the full detail.

Tip

Picking just one? As of June 2026, GPT models still edge out Claude on day-to-day research coding. If your budget only fits one subscription, start with ChatGPT Plus/Pro and Codex CLI. Claude Code is excellent and the two complement each other well when run together — but a one-tool budget still points at OpenAI today.

Important

After Step 1, ask the AI. Once claude or codex is running, paste your intent (e.g. "install stata-mcp and wire it up") and let the AI handle the commands. Sub-pages still document the raw commands, but the recommended path is to use the AI as your installer and configurator.

Contents

1. Install
2. Configure & use effectively
3. Typical research workflow
4. Work with research software
5. Backup & cloud

---

Step 1 — Install

Goal: codex and claude running in your terminal.

This step requires running commands yourself — the AI can't install itself yet. The official installers below are self-contained — no Node.js required.

Windows (PowerShell):

irm https://chatgpt.com/codex/install.ps1 | iex   # Codex CLI
irm https://claude.ai/install.ps1 | iex           # Claude Code

macOS:

curl -fsSL https://chatgpt.com/codex/install.sh | sh     # Codex CLI
curl -fsSL https://claude.ai/install.sh | bash           # Claude Code

Verify with codex --version and claude --version. The first run of either tool opens a browser to authenticate against your ChatGPT or Claude account. If a command isn't found or Codex exits with no output, see troubleshooting (on some bare Windows installs Codex needs the Visual C++ runtime).

→ Details: docs/install.md — Homebrew/npm alternatives, prerequisites, troubleshooting.

Next: Step 2 — Configure

---

Step 2 — Configure & use effectively

Goal: Both tools running at maximum capability.

Ask Codex or Claude Code to configure itself:

"Configure yourself for research work: use the latest model, set reasoning effort to xhigh, keep a sensible permission mode, and add a sensible statusline. Show me the diff before applying."

What that actually opts you into:

• Latest model + deep reasoning. Codex: gpt-5.5 + xhigh. Claude Code: claude-opus-4-8 (alias opus) + effortLevel: "xhigh". Effort runs low | medium | high | xhigh | max; persist xhigh (Opus 4.8/4.7 only) as your research default — the model itself defaults to high — while max is the deepest, session-only level.
• Maximum thoroughness on demand. For exhaustive, correctness-critical work, turn on ultracode — xhigh plus automatic multi-agent workflows: /effort ultracode per session, or make it your default by launching with claude --settings '{"ultracode": true}'. Needs Dynamic Workflows enabled (on by default for Max/Team/Enterprise; on Pro, flip them on in /config). Reach for it on complex or large coding projects; for quick, interactive iteration use /fast (the same Opus, ~2.5× faster output) or a lower effort level.
• Auto mode for execution. Claude Code's auto permission mode runs lower-risk tool calls automatically (after a risk and prompt-injection check) and blocks the rest. It's opt-in, not the out-of-box default — enable it and make it your default with "defaultMode": "auto" in user settings. Codex: on-request (or never per-project). Drop to plan (read-only) or acceptEdits when you want a tighter leash.
• One keyboard rule for Codex. Enter sends now (or injects mid-turn). Tab queues for the next turn.

→ Details: docs/configuration.md — full settings.json / config.toml reference, statusline scripts, permission modes deep-dive.

Next: Step 3 — Workflow

---

Step 3 — Typical research workflow

Goal: A daily flow that doesn't fight you.

1. Open the terminal at your project folder.

• Windows: in File Explorer, navigate to the folder. Type cmd (or powershell, or wt) in the address bar and press Enter.
• Mac: in Finder, right-click the folder → Services → New Terminal at Folder. Or cd from an open terminal.

2. Run a 4-pane window with an AI agent in each pane — each working on its own project or task in parallel:

┌──────────────────┬──────────────────┐
│  AI agent 1      │  AI agent 2      │
│  Project A       │  Project B       │
├──────────────────┼──────────────────┤
│  AI agent 3      │  AI agent 4      │
│  Project C       │  Task X          │
└──────────────────┴──────────────────┘

Snap four windows to quarters. Windows: Win+Left/Right then Win+Up/Down. Mac: install Rectangle for quarter-tiling shortcuts. Or keep all four in one terminal window with internal panes (Windows Terminal: Alt+Shift+D; iTerm2: Cmd+D / Cmd+Shift+D).

3. Resume yesterday's session — claude -c or codex resume --last so context carries over.

Tip

Ask the AI to double-check after it finishes. When the AI completes any complex task, send a brief follow-up: "Double-check everything is correct. Zero errors. Improve. Make everything professional." It re-examines its own work — catches mistakes, polishes weak spots. One of the cheapest accuracy boosts available. For code changes specifically, run /code-review (or /code-review ultra for a deep multi-agent cloud review of the branch or a PR).

→ Details: docs/workflow.md — pane shortcuts, parallel agents, daily flow.

Next: Step 4 — Software

---

Step 4 — Work with research software

Goal: AI that can drive your econ stack.

General pattern: paste a tool's GitHub URL into Claude Code or Codex and ask "install this." For research-specific tools:

Software	What to paste
Python, MATLAB, R, Julia	Native — no install needed. Just describe the task.
Stata	"Install stata-mcp from https://github.com/hanlulong/stata-mcp and wire it up as an MCP server so you can run .do files."
Academic writing	"Install the econ-writing-skill from https://github.com/hanlulong/econ-writing-skill."
Overleaf	Use Overleaf's Dropbox sync. If the 10–20-minute sync lag bites: "Install overleaf-sync-now from https://github.com/hanlulong/overleaf-sync-now."
Anything else	Browse awesome-ai-for-economists, pick a URL, ask the AI to install.

→ Details: docs/software.md — what each tool does, prereqs, the raw install commands.

Next: Step 5 — Backup

---

Step 5 — Backup & cloud

Goal: Never lose work. Code versioned, text canonical, both synced.

Recommended dual-cloud setup:

• Code → Dropbox + GitHub. Dropbox is your live working folder (auto-syncs across machines). GitHub is your versioned backup (history, branches, sharing). Push regularly.
• Text → Dropbox + Overleaf. Edit .tex files locally in ~/Dropbox/Apps/Overleaf/<project>/. Overleaf is the canonical render and what coauthors see.

No Dropbox subscription? OneDrive works as a substitute for the sync layer.

Tell your AI the folder convention. Add this to your project's CLAUDE.md / AGENTS.md:

Use ~/Dropbox/Code/<project>/ for code. Copy final outputs (tables, figures, CSV files) to ~/Dropbox/Apps/Overleaf/<project>/ and edit the .tex files directly there.

For a new project, ask the AI:

"Create a private GitHub repo for this project, set the remote, and push the initial commit."

→ Details: docs/github.md — git + gh install/auth (interactive, do this once), what to back up vs. not, the full dual-cloud workflow.

---

License

MIT