Skip to content

Let's Work on the Next Task: Claude Code, GitHub, and the Most Diligent Project Manager I've Ever Had

Author/Source: Panos Ipeirotis, Behind the Enemy Lines blog, published 2026-03-04

Key Ideas

  • The key insight is that "the repo is the context": a GitHub repository with CLAUDE.md (project master plan) and TASKS.md (to-do list) gives Claude Code persistent memory across sessions, eliminating the need to re-explain the project every time.
  • Claude Code on the Web connects to GitHub repositories, does real work (writing code, documents, analyses), and submits changes as pull requests for human review -- no coding knowledge required.
  • Pull requests are "redlined documents for coders" -- the same tracked-changes workflow familiar from Word or legal documents, applied to any text-based knowledge work.
  • A real example: the author uploaded a CSV dataset and Claude replicated a full experimental analysis (Python scripts, figures, tables with bootstrap CIs, and a rewritten paper section) in under an hour. Two hours total including review.
  • Additional context files (SCHEDULE.md, DECISIONS.md, STYLEGUIDE.md, CHANGELOG.md, BLOCKERS.md, etc.) add structure as projects grow, simulating project management tools with plain markdown.
  • The workflow applies to any text-based knowledge work: research papers, reports, course materials, grant proposals -- not just software.
  • The author acknowledges this is "duct-tape" infrastructure: markdown files parsed by AI are a hack, but it works for current tools and team sizes.

Summary

This article is a step-by-step guide for non-programmers to set up an agentic AI workflow using Claude Code on the Web and GitHub. The central argument is that the chat-based AI interaction model is fundamentally limited because every new conversation starts from zero. GitHub solves this by serving as persistent context: the repository structure, project notes, task lists, and decision records are all visible to Claude Code at the start of every session. The result is that the prompt for most interactions becomes as simple as "Let's work on the next most important task."

The author walks through a five-step setup: creating a GitHub account and repository, connecting it to Claude Code, letting Claude set up the project structure and CLAUDE.md, creating a TASKS.md file, and configuring a GitHub personal access token so Claude handles all git operations. The article explains pull requests as a review mechanism analogous to tracked changes in Word, making the concept accessible to non-technical readers.

A detailed real-world example demonstrates the power of the workflow: given a dataset and instructions to replicate an existing analysis methodology for a new section of a paper, Claude produced Python scripts, figures, tables, and LaTeX text in under an hour -- work that would have taken a PhD student several days. The author emphasizes that errors in this workflow are not "hallucinations" but the same types of mistakes a careful junior colleague might make: misreading a specification or missing an edge case.

Relevance to Economics Research

This article is highly relevant to economics researchers because it demonstrates how to use AI as a project manager and research assistant for academic paper production. The workflow -- structured repositories, task lists, pull request reviews -- maps directly onto the research paper lifecycle. The example of replicating an experimental analysis from a new dataset mirrors common tasks in empirical economics. The emphasis on reproducibility (every figure generated by a script, every number traceable to code) aligns with growing expectations in the discipline. The markdown-file approach to project management, while admittedly a hack, is practical for small research teams today.