AI Tools & Resources
for PhD Students

Qualtrics MCP Server — AI-powered survey building

GitHub Tweet thread

What it is. A Model Context Protocol (MCP) server that lets Claude control Qualtrics through natural language. 53 tools across 8 domains: survey creation, question design, flow logic, randomization, response export, contact and distribution management, and webhooks.

How it works. Install the MCP server locally (Node.js 22+), connect it to your Qualtrics API token, and Claude can build and modify surveys via conversation. You can say things like “Create a 2×2 between-subjects survey with attention checks and randomized block order,” and it generates the full instrument.

Survey experiments are a core method for many of you — especially in BDM, marketing, and OB. The tool collapses the time from experimental design to working instrument. Particularly valuable for rapid prototyping, complex randomization and branching logic (which is error-prone manually), and iterating on designs after advisor feedback without rebuilding from scratch.

Built by Yamil R. Velez (political scientist; has been working on AI-generated experimental designs since 2022).

Aniketa Panjwani — free 3-hour Claude Code course for complete beginners

Tweet thread

A free video course assuming zero prior experience with Claude Code. Walks through Skills (including a live build of a Gmail email-triage skill), MCPs (live exercise), and the actual workflow — not just the marketing. A second-layer alternative to PGP for students with terminal experience who want a denser walkthrough.

Pair this with Backman’s “Claude Code in VS Code” guide if you have the bandwidth.

Aniketa Panjwani — free 4+ hour Codex Desktop App course (in progress)

Tweet thread

Recording in progress as of late April 2026. Notable for being Codex-focused rather than Claude Code. Useful side-by-side. From her training of 20 economists: Sonnet 4.6 on medium thinking did “a pretty good job of converting a paper to Beamer slides for an upcoming NBER talk.”

Many of you will end up on the OpenAI side because of cheap institutional access; worth knowing the Codex workflow even if you default to Claude.

Chris Blattman — claudeblattman.com

claudeblattman.com Tweet thread

A non-coder learning path written by an academic (UChicago Harris): chatbots → prompt engineering → AI project folders → Claude Code install → CLAUDE.md → Skill Library → workflow patterns. Updated April 2026.

Aimed explicitly at non-coders. A good complement to PGP for students from accounting, marketing, and strategy backgrounds who want a written rather than video-first on-ramp.

Anthropic — “Coding Agents” head, 30-minute talk

Tweet pointer

A 30-minute talk by the head of Anthropic’s “Coding Agents” research effort. Covers the philosophy of why and how agentic coding works, not just the mechanics.

For students who want to understand the system rather than just operate it — especially valuable before designing your own Skills or agents.

Karpathy — “LLM Knowledge Bases”

Tweet thread

Karpathy’s gist on using LLMs to build personal knowledge bases for research topics. The idea: dump everything into a folder, let an LLM organize, build cross-referenced markdown wikis you can query.

A natural fit for the lit-review phase of a PhD. You can build a persistent knowledge base for your dissertation topic that grows with every paper you read.

Mihail Velikov — LLM knowledge base for using AI in business research

Tweet thread

Velikov spent a few hours with Claude Code building a knowledge base on AI in business research, inspired by Karpathy’s post.

Concrete academic application of the Karpathy pattern. Fork it for your own field.

“Claude + Obsidian” setup

Tweet thread

A daily-driver setup combining Claude Code with Obsidian for personal knowledge management. The morning workflow has Claude already aware of who you are, what you’re working on, and what’s on your plate.

For students who already use Obsidian (more common in BDM and marketing students than you’d expect). Obsidian-as-RAG is a strong alternative to building a custom MCP server.

Claude Design — slides and prototypes from chat

Announcement Joshua Gans’s reaction Tips thread

A research-preview Anthropic Labs product (Pro / Max / Team / Enterprise) that turns natural-language conversation into prototypes, slides, and one-pagers, powered by Opus 4.7’s vision model.

Gans: “Holy crap. I fed Claude Design a paper and in like 15 minutes it produced slides for a presentation that are amazing. Much better than NotebookLM.”

For seminar talks, job-market practice talks, and research-meeting one-pagers. Worth comparing head-to-head with NotebookLM — the consensus is that Design is meaningfully better for slides.

Sandro Ambuehl — beamer-style slides in HTML, with AI

Tweet thread

Argues PDF/Beamer has inherent limitations and that AI now makes it easy to produce HTML slides that render LaTeX equations, support animations, custom TOCs, and progressive disclosure.

Worth knowing about as an alternative for job-market and conference talks where animation and interactivity actually help. Not a Beamer replacement, but a useful tool to have.

Claes Backman — Claude Skills for paper feedback and code-vs-paper checking

Tweet thread

Two free Claude Skills. One generates feedback on your own academic papers. One compares the code in the replication package to what is claimed in the paper.

The code-vs-paper checker is genuinely novel — replication-style auditing of your own work before submission. The feedback skill is a useful “mock referee.”

Open-source AI paper reviewer

Tweet thread

A free, not-for-profit AI paper reviewer that the author claims rivals reviewer3.com, RefineInk, and Stanford Agentic Reviewer. Costs <$2 per paper via OpenRouter.

Cheap mock-referee feedback before submission. Pair with Backman’s skill for two independent passes.

Anup Malani — train an AI critic on your own edits

Tweet thread

Malani says only 22% of his AI first-drafts were initially usable; after he added a critic agent trained on his own historical edits, that number rose substantially. The fix is a writing-pipeline architecture, not a prompt.

A clean example of using sub-agent / Skills patterns for a research-writing workflow. The “team of critics” idea generalizes to dissertation writing.

Mushtaq Bilal — generate “100% human-written” text with Claude Code

Tweet thread

A workflow for using Claude Code to assist writing without producing the AI-stylistic tells (em-dashes, “delve,” “intricate,” and so on).

Increasingly relevant given growing detection and reputational hazards. Read it critically — the underlying lesson is about editing pipeline, not magic.

Felix Chopra & Ingar Haaland — AI-driven qualitative interviews

Tweet thread

Working paper using LLMs to conduct (and code) qualitative interviews at scale. The argument: qualitative interviews offer a richness quantitative work cannot match but are rarely used in economics — AI lowers the cost.

For students interested in mechanism-based or behavioral fieldwork, this is a real method extension, not a productivity tool. Pairs with our discussion of belief elicitation in L05.

GPT for psychological text analysis (Van Bavel et al.)

Tweet pointer

Tests GPT-3.5 and GPT-4 on detecting psychological constructs (e.g., emotions) in text across 12 languages. Beats dictionary methods substantially.

A useful baseline for any text-as-data work — earnings calls, FOMC minutes, news, social media — relevant across accounting, finance, marketing, and strategy.

Manning & Horton — General Social Agents

working paper · in syllabus

Featured in Lecture 6, Part V-A. Theory-grounded LLM agents predict human behavior in novel games out-of-sample, beating Nash and cognitive-hierarchy benchmarks.

The methodological frontier the in-class hands-on exercise builds on. If you want to use this pattern in your own research, read both Manning & Horton (theory) and DellaVigna & Pope (the prediction tournament).

Ethan Mollick — AI matches expert economists on the same dataset

Tweet thread

A re-run of the famous “146 economist teams, same dataset, wildly different answers” study using agentic AI. Claude Code and Codex land near the human median, but with much tighter dispersion and no extreme answers.

Talking point for whether AI is a substitute or a complement for empirical work. The honest answer is regularizer — it compresses dispersion. Useful for advisor conversations about replication and robustness culture.

Caitlin Knowles Myers — what Claude can build with little prompting

Tweet pointer

Live demo on the “Mixtape: Odd Couple” podcast (with Scott Cunningham) of Claude generating something research-grade from minimal prompting.

A first-person economist demo (rather than the usual SV evangelism) — good gut-check.

Isaiah Andrews — advice for econ PhD students using AI

Emily Oster sharing

Isaiah Andrews’s written advice for econ PhDs on using AI. Oster: “This should probably be circulated to all PhD cohorts.”

The single most balanced piece of advice on AI use in an econ PhD I’ve seen. Worth pairing with our class discussion.

alz_zyd_ — “embrace the chaos” advice for econ/finance PhDs

Tweet thread

Argues that AI will disrupt existing field hierarchies and that the right strategy for a PhD student is to lean into the disruption rather than defend the old equilibrium.

Less practical, more strategic. A useful provocation alongside Andrews’s measured advice.

Bakker, Liu, Christian, Dumitrache — “boiling the frog” RCT on AI use

Tweet thread

A series of RCTs showing that after just 10 minutes of AI assistance, participants perform worse and give up more often than those who never used AI. Cumulative dependence forms quickly.

The behavioral cost of AI use is a research question, not a moral panic. Worth flagging if you find yourself uncritically all-in.

alexolegimas — ML/AI in the standard graduate curriculum

Tweet thread

Argues ML/AI methods will soon be necessary for almost all empirical/analytical work, and pushes back on “can’t you just let Claude Code rip” with practical caveats.

Nuanced take on whether to learn ML methods explicitly or just learn to direct AI. Honest answer: both, with an emphasis on knowing when to do each.