Claude Code as Research Infrastructure — From Chatbot to AI Research Team

Stop using AI as a chatbot. Start building it into your research infrastructure. A presentation from a researcher showed how Claude Code transforms from a conversation partner into a persistent, multi-agent research system: CLAUDE.md as project memory, Skills as domain expert specializations, and parallel AI agent teams that turn a single researcher into the coordinator of an AI intelligence squad.

The Three Layers

Layer 3: AI Intelligence Team
  Multiple agents working in parallel
  Researcher = coordinator, not executor
         ↑
Layer 2: Domain Expert Personas
  Skills, Commands, Personas
  "Reviewer #2", "Editor", "Stats Checker"
         ↑
Layer 1: Persistent Project Memory
  CLAUDE.md = project manual
  Goals, decisions, data locations, code standards

Layer 1: CLAUDE.md as Research Memory

Every time Claude Code starts, it reads CLAUDE.md — your project’s persistent, editable memory:

Without CLAUDE.md, AI is a forgetful goldfish — brilliant but amnesiac. With it, AI becomes a colleague who remembers everything about your project.

Layer 2: Domain Expert Personas

Through Skills, Commands, and Personas, you can weaponize Claude as different specialists:

You’re not asking Claude to “be” these people — you’re loading specific instruction sets that activate different reasoning modes and quality standards.

Layer 3: AI Intelligence Team

The most powerful concept: launch multiple independent AI agents simultaneously, each handling a different research task:

                 ┌─────────────┐
                 │ Researcher  │
                 │(Coordinator)│
                 └──────┬──────┘
          ┌──────┬──────┼──────┬──────┐
          ▼      ▼      ▼      ▼      ▼
       ┌─────┐┌─────┐┌─────┐┌─────┐┌─────┐
       │Code ││Data ││Stats││Lit  ││Result│
       │Audit││ Viz ││Check││Review││Report│
       └─────┘└─────┘└─────┘└─────┘└─────┘
          │      │      │      │      │
          └──────┴──────┼──────┴──────┘
                        ▼
                 Consolidated Output

Each agent works in its own context (isolated worktree), produces results, and the researcher consolidates. Humans shift from executors to managers of AI teams.

The Research Workflow Revolution

The Uncomfortable Question

The article ends with a provocation:

“If experimental validation costs, data visualization costs, and literature review costs all approach zero — how much meaning is left in our 8-hour daily ‘brick-laying’?”

This isn’t about replacement. It’s about reallocation: when AI handles the mechanical research tasks, what do researchers spend their freed time on? The answer should be: thinking harder, asking better questions, and doing the creative work that AI can’t.

Case Study: Building This for Real

I actually built this 3-layer infrastructure for my own research — a formal verification project (relational cell morphing with CHC constraint solving). Here’s exactly what I set up and how it works in practice.

What I Built

Layer 1 — Research CLAUDE.md Template

Instead of writing a CLAUDE.md from scratch for each project, I created a reusable template at ~/.claude/templates/research-CLAUDE.md with pre-structured sections:

# Research Project: 
> PI: ... | Started: ... | Status: ...

## Project Goal         ← 1 sentence
## Research Questions   ← numbered RQs
## Key Hypotheses       ← with predictions
## Decisions Log        ← date | decision | rationale
## Data & Artifacts     ← what | location | format
## Code Standards       ← language, build, test commands
## Current Progress     ← done / in progress / next
## Open Questions       ← things that need answering
## Key References       ← with why-it-matters annotations
## Persona Activation   ← which commands to use
## Parallel Agent Workflow ← copy-paste recipes

For my cell morphing project, the CLAUDE.md includes solver-specific instructions: “always use pcsat_tbq_ar.json config”, “never use integer division in goal constraints”, “run non-vacuity check after every sat”. These rules get enforced automatically across every session.

Layer 2 — Five Persona Commands

I created these as global slash commands (in ~/.claude/commands/), so they work in any project:

Each command is ~80 lines of specific instructions — not a vague “be a reviewer”, but a checklist-driven workflow that produces consistent, structured output.

Layer 3 — Parallel Agent Recipes

Built into the CLAUDE.md as copy-paste workflows:

## Pre-submission Review Sprint
Agent 1: /reviewer2 paper/        — hostile review
Agent 2: /stats-checker paper/    — validate methods
Agent 3: /editor paper/intro.tex  — polish introduction

## Literature + Writing Sprint
Agent 1: /lit-scout "relational cost verification CHC"
Agent 2: /editor paper/related.tex
Agent 3: /viz-designer "benchmark comparison chart"

The Bootstrap Command

The key insight: make setup frictionless. I created /research-setup — a command you run inside any existing project folder:

cd ~/my-research-project
# Then in Claude Code:
/research-setup

It scans the directory, infers the project type from existing files (.clp → formal methods, .py → Python, .tex → paper), and adds only what’s missing. Never overwrites, never creates unwanted directories, asks before changing anything.

What Changed in Practice

The biggest win isn’t any single persona — it’s the compound effect. When CLAUDE.md remembers your project state, personas know how to review your specific domain, and you can run three agents while thinking about the next experiment — you’re operating at a fundamentally different pace.

How LearnAI Team Could Use This

• Adapt this model into a reusable research and documentation infrastructure: CLAUDE.md files for project memory, persona commands for review/editing/literature work.
• Run parallel sessions for audits, synthesis, and publishing workflows.
• Bootstrap new research projects with repeatable templates and commands.

Real-World Use Cases

• Maintain project-specific CLAUDE.md files for LearnAI research, docs, and course repos.
• Use reviewer/editor/lit-scout personas to improve articles before publication.
• Run parallel Claude Code sessions for code audit, source verification, and summary drafting.

Getting Started

1. Create a CLAUDE.md for your research project — or use /research-setup to bootstrap one from a template
2. Build 2-3 personas — start with “Reviewer #2” and “Editor” as slash commands in ~/.claude/commands/
3. Try one parallel session — run a literature search agent while you work on analysis
4. Iterate — update CLAUDE.md as your project evolves
5. Make it reusable — create a /research-setup command so every new project gets the infrastructure automatically

Resources: ClaudeCodeTools Presentation (PDF)