AI Research Workflows
A practice library for applied economists using coding agents for implementation, verification, and research documentation.
This library is the workflow layer behind Too Early To Say. It shows how an applied economist can use a coding agent from project setup through verification and handoff. The agent handles file search, first-pass implementation, reruns, and routine documentation. The estimand, identification, and release decision remain ours.
This is for economists and policy researchers who already know their methods and want a coding agent inside the workflow. The problem it solves is simple. Time goes to implementation and coordination, while design, diagnostics, and interpretation compete for what remains.
One idea runs through the library. A workflow claim needs an experiment. The linked articles therefore separate a teaching pattern from a measured result and state when a numerical case is article-reported rather than publicly reproduced.
[1] Anthropic. "Claude Code overview." Accessed July 21, 2026.
[2] Anthropic. "How Claude remembers your project." Accessed July 21, 2026.
| Capability | Local coding agent with permission | Standalone chat without connected tools |
|---|---|---|
| Reads project files | Can inspect permitted local files | Uses text and files supplied to the session |
| Writes and edits files | Can edit permitted files in place | Produces text or files for review and transfer |
| Persistent context | Can load repository instructions such as CLAUDE.md [2] | Depends on the product's project, memory, and connector settings |
| Runs shell commands | Can run approved commands and tests | Only when the chat product provides a code-execution tool |
| External data sources | Available through configured connectors, APIs, or MCP servers | Available when search or connectors are enabled |
| Custom automation | Repository instructions, reusable skills, hooks, and delegated tasks | Varies by product and plan |
| Research advantage | The implementation can be inspected, tested, and rerun in the project environment | Useful for bounded reasoning and drafting when the required context is supplied |
Workflow guides
Getting Started
A First Session with Claude Code
A practical walkthrough of what Claude Code can and cannot do, with prompting patterns and a complete first-task example.
A Starter Kit for the Economist's First Week in Claude Code
A starter kit for economists using Claude Code: one CLAUDE.md template, a verification checklist, three starter skills, and a week-one glossary.
The kit's seven files are cloneable from the companion repository on GitHub.
Claude Code Guide: From First Session to Personal AI Infrastructure
A comprehensive 14-article guide to mastering Claude Code. Context management, session workflows, agent spawning, and building personal AI infrastructure.
Why It Forgot Everything: Understanding Context
Understanding how AI context windows work, why sessions reset, and how to work with this fundamental limitation of large language models.
Claude Code Guide Series
Creating Skills for Research
Skills are recipe cards for research tasks. Write the steps once, save them in a file, and Claude Code follows those instructions whenever needed.
Creating Helpers: When to Delegate Work
When to create separate Claude Code helpers for focused work, how to design tasks that are easy to hand off, and patterns for running multiple helpers at once.
Building Our Research System: Putting It All Together
How CLAUDE.md, skills, hooks, and MCP servers coordinate project instructions, task automation, and data connections in one workspace.
Hooks: Automation Without Asking
Hooks are automatic triggers that run without asking, like auto-save but for research tasks. A power-user feature, entirely optional.
Connecting Claude to Outside Services: FRED, Census, and Beyond
How to connect Claude Code to external data sources like FRED, Census, and Google Scholar, bringing integrated research workflows into natural conversation.
Research Workflow
What We Mistake for AI Capability
A conceptual hypothesis about how model capability, specification precision, and task tolerance jointly shape perceived output quality.
What Agents Actually Do (And What They Don't)
An agent is a specification-bounded process. Its output quality depends on prompt precision and project context, not hidden model capabilities.
One Context File: A Workflow for Persistent Project Context
A teaching workflow for documenting project conventions in CLAUDE.md. Its session times, prevented-error counts, and savings estimates are not independently timed.
Methods-to-Code with AI: An Article-Reported Workflow
A precise methods paragraph serves as an implementation specification. The transit example and timing comparison are article-reported and not publicly reproduced.
Research Phases Need Different Prompts
Exploration, implementation, and documentation require different AI prompting strategies. Match the prompt to the phase.
Question-First Data Tagging: Finding Forgotten Datasets
Tag data by the questions it can answer, not just what it contains. Question-first tagging records answerable questions so a researcher can retrieve relevant files later.
Code Management
Cleaning a Research Codebase: An Article-Reported Example
A teaching workflow for mapping dependencies and reorganizing code. The script counts, changed imports, hashes, and timing claims are not publicly reproduced.
Reading Our Analysis Files: How Claude Sees Our Research Code
How Claude Code explores research projects using three core tools: Read (look at a file), Glob (find files by pattern), and Grep (search inside files).
Copy-Paste vs. Agent Coding: An Illustrative Comparison
A teaching comparison of chat-based and repository-aware workflows. The cycle count and error story are illustrative scenarios, not measured benchmarks.
The Verification Tax: Every AI Output Needs Checking
A checking workflow for catching errors before they compound across a research project.
Advanced Topics
Context Window Budgeting: Treating Tokens as a Finite Resource
Treating tokens as a finite resource, and knowing when to spawn agents versus work directly.
The Cold Start Problem: Why the First Five Minutes Matter Most
A workflow for loading project context at the start of an AI session and reducing avoidable setup work.
End-of-Session Hygiene: What to Capture Before Context Resets
What to capture before context resets so the next session can recover the project state.
Monitoring Government Data Portals
A case study in tracking California health data releases with Claude Code. Detect new data releases without manual checking.
Building a Literature Surveillance Skill
Automating academic paper discovery with Claude Code. Run one command to query SSRN, NBER, and Google Scholar on a schedule.
Staging LinkedIn Posts with Browser Automation
A case study in form-filling workflows that keep humans in the loop. Browser automation handles navigation while the human retains final approval.
Fix AI Data Visualization: Why Claude Fails (+ Solution)
Structured prompts make layout and verification constraints explicit. The article's numerical figure examples are hypothetical teaching inputs.
Reading Claude Code Usage Data
How to interpret the Claude Code /insights report at beginner and intermediate levels. Same data, different lessons.
AI Econometrics: Using AI for Code, Not for Identification
A conceptual division of labor for AI-assisted econometrics. The exact outputs in its worked examples are article-reported and not publicly reproduced.
How do we know an AI's estimator does what we meant?
AI-generated econometric code can run without error and still be wrong. A routine to verify it: spec the low-visibility choices, plant a known truth, and read the code against its source.
How to tell whether a double machine learning estimate is right
Double machine learning in Python: why a naive plug-in estimates 0.55 when the planted effect is 1.0, how cross-fitting estimates 0.97, and the confounder it still cannot detect.
Running Claude Code skills, for applied economists
A setup guide for the public TETS skill skeletons. Some workflows require user-supplied inputs and do not ship with a turnkey example run.
What AI Impact Looks Like in the Slow Data
AI use telemetry like the Anthropic Economic Index has limits. To measure displaced workers and non-users, we need parallel infrastructure for slow public data.
Key takeaways
- Implementation speed is an empirical claim. The linked timing examples state when they are article-reported rather than backed by a public run record.
- The session loads project instructions from CLAUDE.md. The linked article teaches the workflow and labels its time-savings estimates as independently unmeasured.
- Every AI output needs checking. The verification tax is real: code compiles and runs but may contain subtle errors that compound across a research project.
- The guides address named workflow failures. Context resets, citation errors, token limits, and incomplete handoffs each get a specific check.
Frequently Asked Questions
What is Claude Code?
Claude Code is a command-line AI tool from Anthropic that can read and write permitted files in a local project. A CLAUDE.md file supplies project instructions when the tool starts a session.
Is programming experience required?
Basic command-line familiarity is helpful. The guides center on configuring Claude Code through context files and workflow design.
What is the verification tax?
The verification tax is the time and effort required to check AI output. Code can run while containing subtle errors. The verification tax article explains how to build checking into the workflow.
What decisions remain with the researcher?
The researcher defines the estimand, defends identification, sets validation tolerances, and decides whether a result is ready to report. The agent assists with search, implementation, reruns, and documentation.