Skip to content

Claude Code

Definition

Claude Code is Anthropic's command-line agentic coding tool that runs in a terminal or inside a code editor (VS Code, Cursor). Unlike browser-based chatbots, Claude Code operates directly on the user's local filesystem -- reading, writing, and executing files -- and can integrate with external services via the Model Context Protocol. It is the primary tool discussed across the majority of sources in this wiki.

Claude Code is installed via npm install -g @anthropic-ai/claude-code and requires a paid Anthropic account (Claude Pro minimum). It operates in three modes: Default (proposes edits, user confirms), Plan Mode (explores without touching files), and Auto-Accept (applies edits autonomously). These modes can be cycled with Shift+Tab, giving the user fine-grained control over how much autonomy the AI has.

Context & Background

Claude Code emerged as a qualitative leap beyond chatbot-era AI tools. Ethan Mollick explains that the capability jump was driven by two advances: models that are far better at autonomous self-correcting work (Opus 4.5/4.6), and an "agentic harness" of tools including compacting (structured note-taking when context fills), skills (on-demand specialized instruction sets), sub-agents (parallel isolated processes), and MCP (third-party integrations). METR data shows that the length of tasks AI can complete autonomously has been increasing exponentially.

As of February 2026, approximately 2.1% of scientists with ORCID-linked GitHub profiles use Claude Code, with economists and social scientists the highest adopters by field (up to 3.4%). Adoption follows a U-shaped curve by seniority: early-career and senior (post-tenure) scientists adopt at higher rates than mid-career researchers, likely reflecting higher risk aversion during the tenure-track period.

Key Perspectives

Blattman (summaries/claude-code-newbies.md) provides the primary onboarding guide for non-technical users. He frames the terminal as the biggest hurdle and systematically addresses it, explaining the three operating modes and structuring setup as a five-step sequence: install Claude Code, optionally install VS Code, create a CLAUDE.md file, and configure MCP integrations. The end result is an AI system with email, calendar, and document integrations that improves over time.

Sant'Anna (summaries/my-claude-code-setup.md) presents the most sophisticated academic setup: a "contractor mode" orchestrator with 10 specialized agents, an adversarial critic-fixer loop (5 rounds, hard role separation), 22 slash commands, and 18 context-aware rules. The template covers the full academic lifecycle (slides, papers, data analysis, replication packages) and has been adopted by 15+ research groups. This represents the ceiling of what a well-configured Claude Code environment can do.

Yang (summaries/how-scientists-use-claude-code.md) provides the only empirical measurement of adoption among scientists. The data shows that economists are the highest adopters, that adoption is still very early (significant first-mover advantage available), and that enterprise access remains a barrier AI labs have not cracked.

Mollick (summaries/claude-code-what-comes-next.md) provides the conceptual framework for non-programmers, explaining compacting, skills, sub-agents, and MCP as the "magic tricks" that make sustained autonomous work possible. His demonstration -- building a functional e-commerce website from a single prompt in 74 minutes -- illustrates the ceiling of autonomous capability.

Van Horn (summaries/claude-code-hacks.md) represents the power-user perspective: 4-6 parallel Claude Code sessions, voice input, bypass permissions, and audio notifications. His core principle -- "plan first, always" -- means every idea immediately becomes a plan.md file before any implementation begins.

Panjwani (summaries/ai-agents-econ-research.md) provides budget-conscious guidance for economists, noting that at $20/month Codex may be preferable to Claude Code, but at $200/month Claude Code's capabilities justify the investment since learning speed is proportional to usage.

Practical Implications

  • Low barrier to entry: Installation requires Node.js and a single npm command. No coding experience is required to use Claude Code productively.
  • Start in Plan Mode: For newcomers, Plan Mode (Shift+Tab) is the safest way to explore -- it reads and proposes without changing anything.
  • Configure a CLAUDE.md: Without persistent configuration, every session starts cold. A CLAUDE.md file gives Claude context about your identity, projects, and preferences from the first interaction.
  • Build incrementally: Start with the basic prompt-plan-review loop, add MCP integrations as needed, and build skills over time. Workflows compound.
  • Use Git: Claude Code is dramatically more powerful with version control. Git provides safety (rollback), memory (commit history), and parallelism (worktrees).
  • Budget for usage: Researchers who invest $200/month report learning curves that are qualitatively steeper than those at $20/month.

Open Questions

  • Will Claude Code remain the dominant agentic coding tool, or will Codex, Cursor, and others converge on similar capabilities?
  • How should universities structure institutional access and token budgets for researchers and graduate students?
  • What is the right balance between Auto-Accept mode (speed) and Default mode (oversight) for different research tasks?
  • How will the 2.1% adoption rate among scientists evolve, and what will drive the transition from early adopters to mainstream?

Sources