Setting Up AI Agents for Academic Research

Module 0: Why Should Researchers Care About AI Agents?

"The Principal Investigator Model"

I'm not a statistician. I'm not a programmer. But I've run research labs, managed graduate students, and coordinated multi-site studies. I know how to identify the right person for the right task, how to evaluate their work, and how to synthesize outputs into publishable research.

Working with AI agents is the same skill set. You don't need to become a programmer. You need to become a sophisticated research manager — someone who knows which agent to deploy for which task, how to read their outputs, how to orchestrate them into rigorous research workflows, and when to spot something that feels off. That gut feeling when an agent's output doesn't quite add up? We call this AI intuition — and it's a skill you'll develop.

The Philosophy Shift

What This Tutorial Teaches

This isn't a Python/R/Linux tutorial. This teaches you to:

• Prompt agents in natural language to accomplish research tasks
• Configure pre-built frameworks rather than coding from scratch
• Validate agent outputs using standardized research methodologies
• Orchestrate multiple agents for multi-author validation studies

The Intuitionist Approach

The Intuitionist methodology is something we're actively developing at AgentAcademy. It applies dual-process cognitive theory to AI-assisted social science research:

1. System 1 thinking (fast, intuitive, creative) — Let AI agents brainstorm freely, generate ideas, explore possibilities
2. System 2 thinking (slow, deliberate, critical) — Use different AI agents to rigorously validate, critique, and refine
3. Document everything with standardized templates for reproducibility

The key insight: AI agents can embody both thinking styles, and by orchestrating them properly, researchers can harness creative exploration AND rigorous validation in the same workflow.

Module 1: The Terminal — Your Co-Working Space

The terminal is a text-based interface to your computer. It's faster, scriptable, and creates a log of everything you've done.

Where to Find the Terminal

The terminal is available on every major operating system:

• macOS: Open "Terminal" from Applications → Utilities, or search with Spotlight (Cmd+Space)
• Windows: Use "Windows Terminal" or "PowerShell". For full Linux compatibility, install WSL (Windows Subsystem for Linux)
• Linux: Usually called "Terminal" or "Console" — accessible from your applications menu

The VPS Option: Your Dedicated Research Server

Prof. Wayne's preferred setup is running a Linux VPS (Virtual Private Server) — a remote computer in the cloud that you rent by the month. Recommended providers: Contabo and Hostinger ($5-15/month). Other options include DigitalOcean, Linode, and Hetzner. (This is not a product placement — just honest recommendations!)

Why use a VPS for AI research?

• Separation: Your research environment is completely segregated from your personal and office laptops. No mixing of work files, no accidental data leaks.
• Persistence: Long-running agent tasks continue even when you close your laptop. Start a 4-hour analysis, go to dinner, come back and check results.
• Consistency: Same environment every time. No "it works on my machine" problems.
• Security: API keys and sensitive data live on the VPS, not on laptops that could be lost or stolen.

# From your laptop, connect to your VPS
ssh root@192.168.1.100

Let's break this down:

• ssh — The command to make a secure connection to a remote computer
• root — Your username on the VPS. "root" means you have full administrator access
• 192.168.1.100 — A sample IP address (yours will be different). Your VPS provider gives you this unique number when you set up your server

# cd means "change directory" — it moves you into a folder
cd ~/research-projects    # Go to the research-projects folder

# Now start Claude Code
claude

You don't need a VPS to start — your local terminal works fine for learning. But as your agent workflows grow, a dedicated research server becomes invaluable.

Essential Commands

A Note on sudo (Super User Do)

Some commands require administrator access — like installing software. When you see "Permission denied," prefix with sudo:

# This might fail with "Permission denied"
apt install tmux

# This works because sudo gives you temporary admin powers
sudo apt install tmux

The tmux Multiplexer: Keeping Your Sessions Alive

Here's a common frustration: You're running a 2-hour analysis on your VPS. You go to dinner, or your WiFi drops, or your laptop goes to sleep. When you reconnect... your session is gone. The analysis stopped when your connection died.

This is why we use tmux. It creates persistent sessions that keep running on your VPS even when your laptop goes to sleep, your internet connection drops, or you go to bed and come back in the morning.

# macOS
brew install tmux

# Ubuntu/Debian
sudo apt update && sudo apt install tmux

# Start a session
tmux new -s research

Module 2: Your Agent Roster

What Does "Agentic AI" Even Mean?

You've probably used ChatGPT or Claude in a browser — you type a question, it types back an answer. That's conversational AI. Agentic AI is different: these are AI systems that can actually do things on your behalf.

An AI agent can:

• Read and write files on your computer
• Run commands in the terminal
• Search the web and summarize findings
• Execute multi-step tasks without you supervising each step

Think of it this way: regular AI is like texting a friend for advice. Agentic AI is like hiring a research assistant who can actually go do the work — pull the data, run the analysis, write the first draft — and come back with results.

Quick Vocabulary

Here are terms you'll encounter throughout this tutorial, explained in plain language:

• Prompt: The message or question you type to an AI agent. Just like emailing a research assistant — you write what you need, they respond.
• Context window: How much text the AI can "remember" in a single conversation. Think of it like working memory. A larger context window means the agent can read longer documents.
• Token: A chunk of text the AI processes. Roughly ¾ of a word. You pay per token, so this matters for cost.
• API key: A password-like code that lets you access an AI service. You keep it secret like a password.
• Open-source vs. Open-weight: Important distinction for AI models:
  • Open-source = Full training code, data recipes, and weights are public. True transparency.
  • Open-weight = Only final model weights released (you can run it), but training code/data are proprietary. Most "open" AI models are actually open-weight.
• LLM (Large Language Model): The AI "brain" that powers agents. Current leaders (May 2026): Claude Opus 4.7, Claude Sonnet 4.6, GPT-5.5, GPT-o3, Gemini 3 Pro, Kimi K2.6, GLM5.1, DeepSeek v4.

Skills — Extending What Agents Can Do

Skills are pre-packaged capabilities you can add to AI agents. Think of them like apps for your smartphone — the phone works fine without them, but skills add specialized functionality.

Examples of skills:

• A code-reviewer skill that knows best practices for reviewing code
• A literature-search skill that knows how to query academic databases
• A CommDAAF skill that enforces our research validation framework
• A citation-formatter skill that outputs proper APA/Chicago citations

How to use skills in Claude Code:

/skill code-reviewer

Or in natural language:

"Use the CommDAAF skill to set up a new validation study for this dataset."

Skills encode domain expertise — the skill author has already figured out the best prompts and validation steps. You're leveraging their expertise without developing it yourself.

The AI Services Landscape

Before we dive into specific tools, here's a map of how AI services are organized:

Mainstream AI Model Providers — Companies that build and host their own AI models:

API Marketplaces — One API key to access many models:

• OpenRouter (openrouter.ai) — The main one we recommend. One account, 300+ models from 60+ providers. Often cheaper because they route to the cheapest provider. Free tier includes 25+ models.
• Ollama Cloud — Focuses on running open-source models. Good for privacy-conscious researchers.

Open-Weight Models from China:

Censorship note: Models hosted in China may filter sensitive topics. Good news: When hosted on US/EU infrastructure (via OpenRouter or Ollama Cloud), they run without censorship — full capabilities for research.

Different Tools for Different Jobs

There are two categories of AI agent tools you should know about:

Terminal-Based Coding Agents — Run in your terminal, can read/write files, execute code:

Messaging-Based RA Managers — Live in Telegram/Discord/Slack as your command center:

The powerful insight: Hermes and OpenClaw aren't doing the heavy lifting — they're your command center. Chat with them on your phone, they take notes, and when you're ready for serious work, they help orchestrate your Claude Code/Codex sessions on your VPS.

When to Use Which

The simple answer: Claude models are the gold standard. Codex (OpenAI) is also excellent. Chinese open-source models (Kimi, GLM, DeepSeek) are significantly cheaper and catching up fast.

Prof. Wayne's Current Stack (May 2026):

• Daily chat companion: Hermes Agent on Telegram, powered by Kimi K2.6 via Ollama Cloud
• Serious research: Claude Code, Codex, Gemini CLI, and OpenCode (powered by Kimi K2.6 and GLM5.1 via Ollama Cloud)

Module 3: Setting Up Claude Code — Your Stellar RA

Why Start with Claude Code?

Claude Code is like having a research assistant sitting at your computer who can:

• Read your files — It can look through your data files and documents
• Write for you — It creates folders, writes scripts, and organizes your work
• Explain as it goes — Unlike most software, it tells you what it's doing and why
• Remember context — It can keep track of long conversations about your research project

What This Will Cost

The recommended approach: Subscribe to Claude Max ($100/month). This gives you generous usage limits sufficient for heavy-lifting agentic research — and predictable costs.

Why not pay-per-use API? API billing can be unpredictable and expensive. A single intensive research session can rack up $20-50 in charges. With Max, you know exactly what you're paying.

We'll show you cheaper alternatives using open-weight models in Module 4.

Installing Claude Code

The simplest way (copy and paste this):

On a Mac or Linux computer, open Terminal and type:

curl -fsSL https://claude.ai/install.sh | bash

On Windows, open PowerShell and type:

irm https://claude.ai/install.ps1 | iex

That's it. When it's done, type claude --version to confirm it worked.

Connecting to Your Account

The easiest approach: just run claude in your terminal and follow the interactive prompts:

claude

The first time you run it, Claude Code detects you're not logged in and guides you through authentication. If you have Claude Pro ($20/month) or Max ($100/month), it will open your browser to log in. Follow the prompts and you're connected.

Your First Conversation

cd ~/Documents
claude

Now you're talking to Claude Code. Try typing:

"Can you create a folder called 'research-project' with subfolders for 'data', 'notes', and 'drafts'?"

You just gave an instruction in plain English, and a computer task got done. That's the whole idea.

Module 4: Token Economics and the Political Economy of AI

The Political Economy Question

AI companies control the means of production. They set prices, decide who gets access, and can change terms at any moment. Recent examples: Anthropic restricting third-party orchestration of Claude Code, OpenAI changing pricing structures.

This isn't doom and gloom — it's just reality. The question: as users, what strategies can we adopt to maximize research output while managing this dependency?

Strategy 1: Understand the Cost Structure

The insight: Chinese open-weight models are 5-10x cheaper than frontier US models — and they're catching up in quality fast.

Strategy 2: Don't Put All Eggs in One Basket

• Claude Code — Best quality, but expensive and locked to Anthropic
• OpenCode — Works with any model provider, gives you freedom
• Hermes/OpenClaw — Can be powered by cheap open-weight models

Smart researchers use expensive models for critical work and cheap models for exploration.

Strategy 3: Use Subscription Plans Wisely

• Max plan ($100/month) — Predictable, sufficient for heavy research
• Pro plan ($20/month) — Good for moderate use
• Pay-per-use API — Unpredictable, use with caution

Module 5: Setting Up OpenCode — Your Multi-Model Freedom

The Key Advantage: You're Not Stuck with One Model

Here's the limitation of Claude Code: you can only use Claude models. Yes, Claude models are arguably the best in the market right now — but they're also expensive. And what if you want to:

• Use GLM5.1 for a quick draft (5x cheaper)?
• Try Kimi K2.6 for a coding task?
• Compare how DeepSeek v4 handles your analysis vs. Claude?

Claude Code can't do any of this. You're locked into Anthropic's ecosystem.

OpenCode solves this. It's built from the ground up to work with multiple model providers:

• OpenCode's own plans — Direct access to various models
• OpenRouter — The marketplace we discussed in Module 4
• Ollama Cloud — Run open-weight models in the cloud
• Direct API connections — Connect to any provider you want

Installing OpenCode

Same pattern as Claude Code:

curl -fsSL https://opencode.ai/install | bash

Then verify it installed:

opencode --version

Configuration — Just Run It

Like Claude Code, the easiest approach is to just run the tool and let it guide you:

opencode

The first time you run it, OpenCode will walk you through:

• Choosing your model provider (OpenRouter, Ollama Cloud, direct API, etc.)
• Entering your API credentials
• Selecting your default model

You can always change these settings later with /settings or /connect.

Why Multi-Model Access Matters for Research

Module 6: Setting Up OpenClaw — AI on Your Phone

Claude Code and OpenCode run on your computer in the terminal. OpenClaw runs on messaging apps you already use — WhatsApp, Telegram, Slack, Discord.

Why does this matter for researchers?

• Work from your phone — Check on your analysis from anywhere
• Get notifications — "Your 500-article coding task finished"
• Collaborate easily — Team members can message the same AI assistant
• Keep working while away — Start a task, go teach a class, come back to results

The Lazy Setup (Prof. Wayne's Approach)

If you already have Claude Code set up, just ask it:

"Look at https://openclaw.ai/ and install it on this server and set up an OpenClaw agent for me to talk to in Telegram."

Yes, this really works. Your existing agent handles the installation.

Manual Setup

npm install -g openclaw@latest
openclaw onboard

The onboarding wizard guides you through setup in under 10 minutes.

Connecting Telegram

OpenClaw needs a way to communicate with you:

1. Create a Telegram bot — Open Telegram, search for @BotFather, use /newbot
2. Get your bot token — BotFather will give you a long token string
3. Enter it in OpenClaw setup — The setup wizard asks for this

Choosing Your AI Model

OpenClaw works with any model provider you prefer:

• Codex Plan — If you have a Codex subscription
• Ollama Cloud — Prof. Wayne uses Kimi K2.6 via Ollama Cloud for his Hermes Agent
• OpenRouter — Access to hundreds of models
• Direct API keys — Connect to Anthropic, OpenAI, etc.

What You Can Do With OpenClaw

Once set up, you can text your AI things like:

• "Review my draft paper and list the main issues"
• "What's the status of my analysis task?"
• "Summarize the 10 most recent papers on [topic]"

Module 7: Keeping Your API Keys Safe

What's an API Key Again?

Remember from Module 2: an API key is like a password that lets you access AI services. You don't want to accidentally share it with the world.

The Danger: Accidentally Sharing Your Keys

Imagine writing your API key directly into a research script, then sharing that script with a colleague or uploading it to GitHub. Now anyone can use your key and charge things to your account. This happens more often than you'd think.

The Solution: Environment Variables

An "environment variable" is just a way to tell your computer: "Remember this value, but don't write it into any files."

Here's the simplest approach:

When you open your terminal, type this (replacing with your actual key):

export ANTHROPIC_API_KEY="sk-ant-api03-your-actual-key-here"

Now start Claude Code:

claude

That's it. Claude Code can now see your key, but the key isn't written in any file that could accidentally get shared.

Making It Permanent

The export command only works for your current terminal session. To make it permanent:

On Mac:

echo 'export ANTHROPIC_API_KEY="your-key-here"' >> ~/.zshrc

On Linux:

echo 'export ANTHROPIC_API_KEY="your-key-here"' >> ~/.bashrc

The Golden Rules

• Never put API keys in files you share — Not in scripts, not in emails, not in screenshots
• Try not to paste API keys into prompts — Use environment variables instead. (Prof. Wayne admits he sometimes pastes keys directly for convenience — it works, but not best practice.)
• If you accidentally share a key, regenerate it — Go to your provider's website and create a new one
• Set spending limits — Both OpenRouter and Anthropic let you cap how much can be charged

Routine Security Audits

Ask your AI agent to audit your own security. Try prompts like:

• "Audit this server for cybersecurity vulnerabilities. Check for exposed API keys, weak permissions, and security misconfigurations."
• "Review my .bashrc, .zshrc, and environment files for any hardcoded secrets."
• "Check if any of my project files contain API keys that might accidentally get committed to git."

Make this a routine practice — once a week or before major project milestones.

Module 8: The CommDAAF Framework

CommDAAF (Computational Multi-Model Data Analysis and Augmentation Framework) is AgentAcademy's quality assurance system for AI-conducted research.

CommDAAF builds on DAAF (Data Analysis and Augmentation Framework), originally developed by the DAAF Contribution Community. CommDAAF extends DAAF for communication research, adding multi-model validation protocols, communication-specific codebook templates, and integration with Intuitionist workflows.

Using CommDAAF as a Skill

CommDAAF is available as an importable skill for Claude Code:

/skill commdaaf

Or in natural language:

"Import the CommDAAF skill and set up a new content analysis study with 3-model validation."

The skill creates proper folder structure, guides codebook definition, sets up multi-model validation, and generates publication-ready methodology documentation.

The Q1-Q3 Validation Framework

Core Principles

• Multi-Model Validation — 3+ AI models independently analyze identical datasets
• Reliability Metrics — Cohen's κ, Fleiss' κ reported for inter-model agreement
• Adversarial Review — AI reviewers critique studies before publication
• Transparent Failures — Corrections and retractions published openly

Module 9: Setting Up Hermes Agent — Your Always-On RA Manager

Both Hermes Agent and OpenClaw (Module 6) live in your messaging apps. The key difference:

• OpenClaw — More technical, designed for researchers who want fine-grained control
• Hermes Agent — Smoother onboarding, stronger memory features, built by Nous Research

Think of them as two options for the same role: your "RA manager" who lives in Telegram/Discord/Slack.

Why a Messaging-Based RA Manager?

Claude Code and OpenCode are powerful — but they run in your terminal. A messaging-based RA manager like Hermes:

• Takes notes for you — "Remember that I'm working on a framing study"
• Stays available 24/7 — Message from your phone while commuting
• Coordinates your agents — "What was I working on yesterday?"

Choosing Your AI Model

Like OpenClaw, Hermes can be powered by any model you prefer:

• Prof. Wayne's setup: Kimi K2.6 via Ollama Cloud — cost-effective, excellent reasoning
• Via OpenRouter: Access to hundreds of models
• Via Codex Plan: If you have an OpenAI subscription

The Lazy Setup (Ask Claude Code)

If you already have Claude Code running:

"Look at the Hermes Agent documentation and install it on this server. Set it up to connect to my Telegram."

Manual Setup

curl -fsSL https://hermes-agent.nousresearch.com/install | bash
hermes setup

Hermes vs. Claude Code: Not Either/Or

Module 10: Research Data Access — The Key Bottleneck

You've learned to set up AI agents. You understand the frameworks. But here's the reality: The biggest bottleneck isn't the AI tools — it's access to research data.

The Data Landscape for Social Scientists

AgentAcademy maintains a curated registry at the Data Landscape tab. Key categories:

Survey Data — Gold Standard Sources:

Political & Policy Data:

Academic & Bibliometric Data:

The Social Media Access Crisis

Working with Your Agent on Data

Once you have data, your agent can help:

"I have a CSV of 10,000 survey responses. Read the file and describe the variables."

"Use the OpenAlex API to find all papers on 'framing' published since 2020."

Module 11: Research Ethics in the Age of AI

AI is transforming academic research faster than institutions can respond. This creates genuine ethical ambiguity.

The Originality Question

If an AI agent writes significant portions of your paper, what does "authorship" mean?

The emerging consensus:

• Authorial responsibility matters, not production method — You stand by the final product regardless of how it was produced
• Disclosure is trending toward standard practice — AI usage declarations should become as ubiquitous as conflict of interest declarations

Data Privacy: When Research Becomes Training Data

When you paste research data into AI prompts, that data potentially becomes part of the AI system's memory.

Practical safeguards:

• De-identify before prompting — Strip names, locations, dates from data
• Use local models for sensitive data — Ollama runs entirely on your machine
• Check provider data policies — API has different retention than subscription

The Training Pipeline Disruption

AI handles much "grunt work" that traditionally trained researchers. As one researcher admitted: "It is rational to adopt these tools. That's precisely why the problems will keep getting worse: no individual has an incentive to slow down."

Practical Ethical Guidelines

1. Disclose AI use — In methods sections, acknowledgments, or author notes
2. Verify agent outputs — Never publish AI-generated claims without checking
3. Protect human subjects — De-identify data before prompting
4. Maintain judgment — You're the PI; final responsibility is yours
5. Don't flood — Use productivity gains to do better work, not more mediocre work

Module 12: Research Ideation

Idea Generator 1: Replication at Scale

Take an existing study (e.g., Entman 1993 on framing) and ask: "Could I replicate this finding at 100x scale using agents to code news archives?"

Idea Generator 2: Multi-Model Reliability

Inspired by Bean et al. (2026): "How reliable are different LLMs at content analysis tasks in my domain?"

Idea Generator 3: Longitudinal Coordination Analysis

Detect coordinated information campaigns: "Are certain actors sharing messages in coordinated patterns?"

Idea Generator 4: Automated Literature Reviews

Synthesize faster with agent assistance: "What are the emerging themes in [topic] over the past 5 years?"

The Study Design Template

1. Research Question
2. Data Source (What exists? What's accessible?)
3. Agent Assignment (Which agent for which task?)
4. Validation Strategy (CommDAAF Q1-Q3)
5. Resource Estimate (Tokens × Cost)

Module 13: Your First Agent-Assisted Research Task

Let's do something concrete. We'll use Claude Code to create a coding scheme for analyzing how news articles frame climate change — a classic communication research task.

Step 1: Create a Project Folder

Open your terminal and type:

cd ~/Documents
mkdir climate-framing-study
cd climate-framing-study
mkdir data output

Step 2: Start Claude Code

claude

Step 3: Ask Claude to Brainstorm Categories

Now you're talking to Claude Code. Type something like:

"I'm doing a content analysis of climate change news coverage. Can you brainstorm 5-7 potential framing categories? For each frame, give me a name and a brief definition."

Step 4: Get a Second Opinion

Here's where it gets interesting. Ask Claude to critique its own work:

"Now pretend you're a different analyst reviewing this list. Which categories overlap? Which need clearer definitions? What's missing?"

Or even better: close this session and use a different AI (GPT-4o via OpenRouter) to review the categories. Different AI = independent validation.

Step 5: Finalize Your Codebook

Based on the brainstorm and critique, ask Claude:

"Create a final codebook with these refined categories. For each category include: name, definition, what counts as an example, and what doesn't count. Save this to output/codebook.md"

What You Just Accomplished

In about 30-45 minutes, you:

1. Generated initial ideas creatively (brainstorming phase)
2. Critically evaluated those ideas (validation phase)
3. Synthesized them into a usable research instrument
4. Documented the process for reproducibility

Notice that you didn't write any code. You had a conversation. The AI did the work. Your job was to know what you wanted, ask clear questions, and evaluate whether the output made sense.

Module 14: How Researchers Work With Agents

You've learned the tools and frameworks. Now let's see how researchers carry out studies with agents — the prompts they use, how they iterate with Claude Code, and how they manage multi-model validation.

Case Study 1: Nonprofit Framing — Autonomous Coding at Scale

Challenge: 465 nonprofit mission statements to code for technocratic framing. Manual coding would take weeks.

Step 1: Project Setup

cd ~/research
mkdir nonprofit-framing-study
cd nonprofit-framing-study
claude

First prompt to Claude Code:

"I'm starting a content analysis study of nonprofit mission statements. I have 465 mission statements from IRS Form 990 filings. I want to code each one for 'technocratic framing.' Help me set up the project structure following CommDAAF conventions."

Step 2: Codebook Development (20 min)

"Based on academic literature on technocratic governance, draft a coding scheme with clear inclusion/exclusion criteria and 3 example mission statements for each category."

After 3-4 iterations refining fuzzy distinctions, the codebook is solid.

Step 3: Pilot Coding (30 min)

"Read the first 20 mission statements and code each one. Output: Organization name, Classification, Key phrases, Confidence level. Format as markdown table."

Researcher reviews, catches errors, refines codebook.

Step 4: Full Autonomous Coding (45 min)

"Now code all 465 mission statements using the finalized codebook. Save results to output/full_coding.csv."

Claude Code processes all 465 statements, reporting progress every 50 items.

Step 5: Validation

"Generate a stratified random sample of 50 for human validation. Calculate inter-rater reliability between AI and my human coding."

Result: κ = 0.78 — acceptable for publication.

Total time: ~3 hours from start to publication-ready dataset.

Read the full study →

Case Study 2: Medical Triage — Multi-Model Testing

Challenge: Test 100 medical scenarios across 5 different models, then have physicians validate.

The Multi-Model Setup (using OpenCode):

opencode

"Test medical triage across 5 models. For each of 100 scenarios, send to: claude-sonnet-4.6, gpt-5.5, gemini-3-pro, kimi-k2.6, glm-5.1. Log each response. Do NOT let any model see other models' responses."

Agreement Analysis:

"Calculate pairwise Cohen's κ between each model pair. Compare to ground truth triage level. Identify high-disagreement scenarios."

The Discovery: Analysis revealed the "safety-burden trade-off" — models that never under-triaged had the highest over-triage rates.

Read the full study →

Case Study 3: Frame Analysis — The Full CommDAAF Protocol

Challenge: 719 social media posts coded by 3 independent models with complete documentation.

Three-Model Setup (separate sessions):

# Terminal 1 — Claude
tmux new -s claude-coding && claude

# Terminal 2 — GLM
tmux new -s glm-coding && opencode --model glm-5.1

# Terminal 3 — Kimi
tmux new -s kimi-coding && opencode --model kimi-k2.6

Identical prompt to all three:

"Code these social media posts for communication frames: (1) Economic, (2) Moral/Values, (3) Conflict, (4) Human Interest, (5) Attribution. Output frame number and one-sentence justification for each."

Key discipline: each model codes in its own session, never seeing other outputs.

Consensus Calculation:

• Unanimous agreement: 605 posts (84.1%)
• Majority agreement: 89 posts (12.4%)
• No consensus: 25 posts (3.5%)

HILAR — Human Review:

"Export all non-unanimous cases for human review. Include each model's classification and reasoning."

Researcher adjudicates each case, documenting reasoning in Q3.

Read the preprint →

The AgentAcademy Workflow Pattern

Prompts That Work

For codebook development:

"Based on [theoretical framework], develop a coding scheme for [construct].
Include: clear definitions, inclusion/exclusion criteria, 3 examples per category."

For autonomous coding:

"Code all items in [file] using the codebook. For each: ID, classification,
key evidence, confidence. Save to [output file]."

For multi-model validation:

"Send same task to multiple models independently. Do not reference other models."

For reliability:

"Calculate inter-coder reliability between [file1] and [file2].
Report: Cohen's κ, percentage agreement, confusion matrix."

Technical Appendix

Terminal Commands Reference

Troubleshooting

Completion Checklist

• Can open and navigate the terminal
• Can create and source environment variables
• Has Claude Code installed and configured
• Can explain the difference between Claude Code, OpenCode, OpenClaw, and Hermes
• Understands token costs and cost routing via OpenRouter
• Can describe System 1 vs System 2 in the Intuitionist Workflow
• Can articulate the CommDAAF Q1-Q3 framework
• Has completed at least one practice research task