The AI setup most journalists don’t know about
A tutorial series on moving from chatbots to a working reporting system
2026 is starting with a clear bet from Anthropic and OpenAI: stop thinking “chatbot,” start thinking “assistant.” Some of it is hype, sure. But the practical upside is real. It’s been a big unlock for me.
I’ve been using this system for a few weeks now, and it’s changed how I work. It worked well enough that I decided to turn it into a short tutorial series, in case it helps other build the same kind of setup.
Module 1: Why Claude Code for Journalists
You’ve probably tried ChatGPT for research. Maybe you’ve asked Claude to help draft an interview question or two. Maybe you’ve abandoned it after a few attempts because it felt like more trouble than it was worth.
But there might be something way more useful than just a few one-offs with chatbots.
When you ping ChatGPT with “help me research this topic” or “draft questions for this interview,” you’re using AI like a magic 8-ball. Ask a question, get an answer, start over. No memory, no context, no system.
What if instead of asking an AI assistant the same questions every morning, you had a system that knew your beat, understood your style, tracked your sources, and could pick up where you left off yesterday?
The Problem: Why Regular AI Chat Isn’t Enough
Let me show you what I mean with a real scenario. Let’s say you’re working on a story about municipal budgets. Over three days, you:
Downloaded 5 budget PDFs from different city departments
Tracked down 8 background articles on previous budget controversies
Collected interview transcripts with 4 council members
Started drafting questions for a follow-up interview with the mayor
In a normal workflow with ChatGPT or standard Claude, here’s what happens:
Day 1: “Can you help me analyze this budget PDF?” → Upload document, get analysis
Day 2: “Can you compare these two budgets?” → Re-upload both documents, explain context again
Day 3: “Based on everything I’ve shared...” → Wait, you haven’t shared anything. This is a new conversation.
Every single time, starting from zero. Re-uploading files. Using a “project” to organize your files for the more advanced users. But still feels like either having an assistant with amnesia or something not really scalable across projects. The overhead of constantly re-establishing context makes it slower than just doing it yourself.
Claude Code vs. Codex: Which One Should You Use?
Quick answer: both are good. The models underneath (Claude Opus 4.5, GPT-5.2) are roughly comparable. The difference is how each company designed the product on top.
Codex (OpenAI) feels more autonomous. When I’m working on technical problems for my startup, Codex tends to be more thorough. It works on its own and delivers complete results. I also appreciate being able to track my remaining quota directly in VS Code. No guessing when you’ll hit limits.
Claude Code (Anthropic) feels more conversational and iterative. That back-and-forth can make it friendlier and easier to use, especially for non-technical workflows. It’s better at being a sparring partner when you’re thinking through problems.
There’s also a buzz factor worth acknowledging. Right now, you’ll see more posts, tutorials, and community momentum around Claude Code. But that could shift quickly (it always does in AI).
Neither is strictly “better” overall. They’re tuned to slightly different working styles. My personal stack: Claude Code for the assistant/research workflows, Codex for coding tasks on my startup. I switch between them depending on what I’m doing.
We’ll explore both in this series. Module 2 covers Claude Code setup, Module 3 covers Codex. The core principles (file access, persistent context, reusable workflows) work with both.
What’s Different About Claude Code (and Cowork)
Wait… Why is this called “Claude Code” if I’m not coding?
Think of it this way. ChatGPT or Claude is like asking a chef for a recipe. Claude Code is like having the chef in your kitchen. They look in your fridge, taste the soup, and adjust as you cook.
The real shift is the interface. Instead of a chat box, you use something like VS Code, a workspace developers use to work with their files. That lets Claude see and work with your actual documents and code instead of you constantly re uploading and re explaining.
It’s not that the model is suddenly smarter. It’s that it finally has hands.
If you want a glimpse of where AI is heading, just look at Anthropic’s Cowork launch on Monday. Cowork is basically Claude Code for non technical users. Same idea of an assistant that can see your files, remember context, and run workflows. Just packaged as an app instead of a developer tool (only in preview right now for Max users).
Here’s what changes:
1. File System Access
Claude can read, write, and edit files directly on your computer. No more uploading the same budget PDF five times. Point Claude to your /sources/municipal-budget/ folder once, and it has access to everything there.
When you get a new document, drop it in the folder and say “analyze the new city budget against last year’s.” Claude knows where to find both files, and you build your own archives gradually.
2. Persistent Context
That municipal budget story? Here’s what you could actually do now:
Create a folder called /stories/municipal-budget-story/ with:
All source documents
Interview transcripts
Research notes
A
context.mdfile that Claude updates with key findings
When you sit down on Day 3, you don’t re-explain anything. Claude has been tracking this story. It knows what we found in the budget, what people said, and what gaps remain.
You just say: “Draft follow-up questions for the mayor focusing on the discrepancies we found in the transit budget.”
Claude already knows what those discrepancies are. We documented them together two days ago.
3. Skills You Can Reuse
Here’s where it gets powerful.
Let’s say you’re a journalist doing FOIA requests regularly. You keep asking Claude to help with the same structure: draft the initial request, track submission dates, draft appeals when denied, analyze responses when they arrive.
Instead of re-explaining this process every time, you could create a skill called foia-request.md in your /.claude/skills/ folder. Now whenever you’re working on a FOIA request, Claude automatically uses that workflow. No re-teaching required.
Same concept for interview prep. You could build an interview-prep.md skill that, when activated, tells Claude to:
Read all background docs in the relevant folder
Generate questions based on what’s already been found
Identify knowledge gaps
Suggest follow-up angles
No programming required. You just write down the workflow you want once, save it as a skill, and it becomes automatic.
4. MCP Integrations (The Advanced Stuff)
This is optional but powerful. MCP (Model Context Protocol) lets Claude connect to external tools and databases.
For example, You could connect Claude to:
Your newsroom’s CMS (to fetch your published articles)
Public records databases
RSS feeds for your beat monitoring
When you ask Claude to “check if we’ve covered this story before,” it searches your published archive. When you say “monitor city council agenda items,” it scans the right sources automatically.
(We’ll cover MCP setup in Module 4. For now, just know it exists.)
Real Workflows: What This Actually Looks Like
Let me walk through three examples, including from my actual work:
Example 1: Daily AI News Monitoring
The old way:
8:00am: Open 20+ tabs (Hacker News, AI Twitter, research papers, company blogs)
30 minutes of FOMO scrolling through AI announcements
Try to figure out what actually matters vs. hype
With your assistant:
8:00am: “/daily-digest”
5 minutes later: Claude gives me a markdown file with 8-10 relevant items, summarized with context
Each item includes: what it is, why it matters, how it connects to previous developments
I can focus on understanding what’s important instead of chasing every announcement
Setup time: 20 minutes to configure which sources to monitor (RSS feeds, specific Twitter accounts, research repos). Now it runs daily and saves 25+ minutes of FOMO scrolling.
Example 2: Interview Preparation
The old way:
Read through scattered notes and articles
Hunt down background, like that YouTube video of the school board meeting
Manually take notes while watching
Draft your questions alone
Wonder if you’re missing angles
With your assistant:
“Fetch the transcript from this YouTube link: [school board meeting]”
Claude grabs the transcript automatically
Drop any other background materials in
/sources/[source-name]/Share your draft questions: “Here are my 5 questions for this interview about school board funding”
Claude acts as sparring partner: suggests follow-up angles based on what was said in the meeting, identifies gaps in your coverage, points to relevant background you might have missed
You refine your approach based on the back-and-forth
Example 3: Publication Package
The old way:
Finish your article draft
Separately write social media posts for Twitter, LinkedIn, Facebook
Draft push notification (character limit headaches)
Spilling coffee on your keyboard (the most important step)
Manually check against style guide
Copy-paste between platforms
Hope you didn’t miss anything
With your assistant:
Finish your article draft in
/stories/drafts/school-funding-story.md“Generate publication package for this article”
Claude creates:
Push notification (60 characters, attention-grabbing)
Twitter thread (5 tweets with key findings)
LinkedIn post (professional tone, context for policy audience)
Facebook post (community-focused angle)
Style guide check flagging any issues
All saved in
/stories/drafts/school-funding-publication-package.mdYou review, adjust, and publish
Here’s where it gets powerful: These three examples actually flow together into one system.
You start your day with the AI news digest (Example 1), which surfaces a relevant story. You use that context to prep your interview questions (Example 2). After the interview, you draft your article. Then you generate the complete publication package (Example 3).
It’s a dumbed-down example, but you get the idea. Claude remembers the full context: what was in the morning digest, what you asked in the interview, what the story is about. You’re not re-explaining at each step. You’re building on a continuous workflow.
That’s the difference between using AI as a chatbot vs. building it into your actual work system.
What it costs
Claude Pro or ChatGPT Plus subscription: $20/month
OR API access: depending on usage
Setup time: 2-4 hours for initial workspace setup
Learning curve: 1-2 weeks to proficiency
Will this save you 10 hours a month? 20? Depends entirely on your workflow and how repetitive your tasks are. Overall, the real benefit for me isn’t so much “how many hours,” but having a persistent assistant.
What This Isn’t
Before we go further, let’s be clear about what this assistant doesn’t do:
It doesn’t write your stories. You’re still the journalist. Claude Code helps with research, organization, and tedious tasks. The reporting, the narrative, the editorial judgment: that’s all you.
It doesn’t replace fact-checking. Claude Code helps structure your fact-checking workflow, but you still verify every claim with primary sources. Never cite Claude’s analysis without independent verification.
You do have to be careful about what you share. You control what Claude can access, and it only sees the folders and files you explicitly give it. That means confidential sources, sensitive documents, or unpublished investigations stay private unless you choose to expose them. We’ll go deeper into how to set this up safely in Module 5.
It doesn’t work magically. You need to set it up properly, build good workflows, and use it consistently. This is a system, not a magic wand. And you can teach it how to improve.
You’re building a system where Claude can actually remember what the hell you talked about yesterday.
That’s the unlock.
What Happens Next
If you’re reading this thinking “okay, this sounds useful but where do I even start,” that’s literally what the next module is for.
Module 2 (in the next few days) walks through the complete setup:
Installing VS Code and Claude Code (30 minutes)
Creating your journalism folder structure
Setting up your first workflow (daily monitoring)
Testing everything with a real task
Module 3 does the same for Codex, so you can compare.
By the end of Module 2, you’ll have a working system, with an actual folder structure on your computer with Claude actively helping with your work.
Then Module 4 covers the advanced stuff (MCP integrations, subagents, deep research workflows), and Module 5 tackles ethics, security, and rolling this out in your newsroom.
One More Thing
I know this sounds like a lot. But remember: you probably already have a system for managing your journalism work. Files in folders, interview transcripts, research notes, source tracking.
All we’re doing is teaching Claude where those files live and what workflows you want it to help with.
You’re not learning to code. You’re just pointing an AI assistant at the work you’re already doing and saying “help with this part. That’s it.
Questions? Drop them in the comments :) See you in Module 2 for the actual setup! 👋
Quick links
ies, in case it helps other build the same kind of setup.
Module 1: Why Claude Code for Journalists
You’ve probably tried ChatGPT for research. Maybe you’ve asked Claude to help draft an interview question or two. Maybe you’ve abandoned it after a few attempts because it felt like more trouble than it was worth.
But there might be something way more useful than just a few one-offs with chatbots.
When you ping ChatGPT with “help me research this topic” or “draft questions for this interview,” you’re using AI like a magic 8-ball. Ask a question, get an answer, start over. No memory, no context, no system.
What if instead of asking an AI assistant the same questions every morning, you had a system that knew your beat, understood your style, tracked your sources, and could pick up where you left off yesterday?
The Problem: Why Regular AI Chat Isn’t Enough
Let me show you what I mean with a real scenario. Let’s say you’re working on a story about municipal budgets. Over three days, you:
Downloaded 5 budget PDFs from different city departments
Tracked down 8 background articles on previous budget controversies
Collected interview transcripts with 4 council members
Started drafting questions for a follow-up interview with the mayor
In a normal workflow with ChatGPT or standard Claude, here’s what happens:
Day 1: “Can you help me analyze this budget PDF?” → Upload document, get analysis
Day 2: “Can you compare these two budgets?” → Re-upload both documents, explain context again
Day 3: “Based on everything I’ve shared...” → Wait, you haven’t shared anything. This is a new conversation.
Every single time, starting from zero. Re-uploading files. Using a “project” to organize your files for the more advanced users. But still feels like either having an assistant with amnesia or something not really scalable across projects. The overhead of constantly re-establishing context makes it slower than just doing it yourself.
Claude Code vs. Codex: Which One Should You Use?
Quick answer: both are good. The models underneath (Claude Opus 4.5, GPT-5.2) are roughly comparable. The difference is how each company designed the product on top.
Codex (OpenAI) feels more autonomous. When I’m working on technical problems for my startup, Codex tends to be more thorough. It works on its own and delivers complete results. I also appreciate being able to track my remaining quota directly in VS Code. No guessing when you’ll hit limits.
Claude Code (Anthropic) feels more conversational and iterative. That back-and-forth can make it friendlier and easier to use, especially for non-technical workflows. It’s better at being a sparring partner when you’re thinking through problems.
There’s also a buzz factor worth acknowledging. Right now, you’ll see more posts, tutorials, and community momentum around Claude Code. But that could shift quickly (it always does in AI).
Neither is strictly “better” overall. They’re tuned to slightly different working styles. My personal stack: Claude Code for the assistant/research workflows, Codex for coding tasks on my startup. I switch between them depending on what I’m doing.
We’ll explore both in this series. Module 2 covers Claude Code setup, Module 3 covers Codex. The core principles (file access, persistent context, reusable workflows) work with both.
What’s Different About Claude Code (and Cowork)
Wait… Why is this called “Claude Code” if I’m not coding?
Think of it this way. ChatGPT or Claude is like asking a chef for a recipe. Claude Code is like having the chef in your kitchen. They look in your fridge, taste the soup, and adjust as you cook.
The real shift is the interface. Instead of a chat box, you use something like VS Code, a workspace developers use to work with their files. That lets Claude see and work with your actual documents and code instead of you constantly re uploading and re explaining.
It’s not that the model is suddenly smarter. It’s that it finally has hands.
If you want a glimpse of where AI is heading, just look at Anthropic’s Cowork launch on Monday. Cowork is basically Claude Code for non technical users. Same idea of an assistant that can see your files, remember context, and run workflows. Just packaged as an app instead of a developer tool (only in preview right now for Max users).
Here’s what changes:
1. File System Access
Claude can read, write, and edit files directly on your computer. No more uploading the same budget PDF five times. Point Claude to your /sources/municipal-budget/ folder once, and it has access to everything there.
When you get a new document, drop it in the folder and say “analyze the new city budget against last year’s.” Claude knows where to find both files, and you build your own archives gradually.
2. Persistent Context
That municipal budget story? Here’s what you could actually do now:
Create a folder called /stories/municipal-budget-story/ with:
All source documents
Interview transcripts
Research notes
A
context.mdfile that Claude updates with key findings
When you sit down on Day 3, you don’t re-explain anything. Claude has been tracking this story. It knows what we found in the budget, what people said, and what gaps remain.
You just say: “Draft follow-up questions for the mayor focusing on the discrepancies we found in the transit budget.”
Claude already knows what those discrepancies are. We documented them together two days ago.
3. Skills You Can Reuse
Here’s where it gets powerful.
Let’s say you’re a journalist doing FOIA requests regularly. You keep asking Claude to help with the same structure: draft the initial request, track submission dates, draft appeals when denied, analyze responses when they arrive.
Instead of re-explaining this process every time, you could create a skill called foia-request.md in your /.claude/skills/ folder. Now whenever you’re working on a FOIA request, Claude automatically uses that workflow. No re-teaching required.
Same concept for interview prep. You could build an interview-prep.md skill that, when activated, tells Claude to:
Read all background docs in the relevant folder
Generate questions based on what’s already been found
Identify knowledge gaps
Suggest follow-up angles
No programming required. You just write down the workflow you want once, save it as a skill, and it becomes automatic.
4. MCP Integrations (The Advanced Stuff)
This is optional but powerful. MCP (Model Context Protocol) lets Claude connect to external tools and databases.
For example, You could connect Claude to:
Your newsroom’s CMS (to fetch your published articles)
Public records databases
RSS feeds for your beat monitoring
When you ask Claude to “check if we’ve covered this story before,” it searches your published archive. When you say “monitor city council agenda items,” it scans the right sources automatically.
(We’ll cover MCP setup in Module 4. For now, just know it exists.)
Real Workflows: What This Actually Looks Like
Let me walk through three examples, including from my actual work:
Example 1: Daily AI News Monitoring
The old way:
8:00am: Open 20+ tabs (Hacker News, AI Twitter, research papers, company blogs)
30 minutes of FOMO scrolling through AI announcements
Try to figure out what actually matters vs. hype
With your assistant:
8:00am: “/daily-digest”
5 minutes later: Claude gives me a markdown file with 8-10 relevant items, summarized with context
Each item includes: what it is, why it matters, how it connects to previous developments
I can focus on understanding what’s important instead of chasing every announcement
Setup time: 20 minutes to configure which sources to monitor (RSS feeds, specific Twitter accounts, research repos). Now it runs daily and saves 25+ minutes of FOMO scrolling.
Example 2: Interview Preparation
The old way:
Read through scattered notes and articles
Hunt down background, like that YouTube video of the school board meeting
Manually take notes while watching
Draft your questions alone
Wonder if you’re missing angles
With your assistant:
“Fetch the transcript from this YouTube link: [school board meeting]”
Claude grabs the transcript automatically
Drop any other background materials in
/sources/[source-name]/Share your draft questions: “Here are my 5 questions for this interview about school board funding”
Claude acts as sparring partner: suggests follow-up angles based on what was said in the meeting, identifies gaps in your coverage, points to relevant background you might have missed
You refine your approach based on the back-and-forth
Example 3: Publication Package
The old way:
Finish your article draft
Separately write social media posts for Twitter, LinkedIn, Facebook
Draft push notification (character limit headaches)
Spilling coffee on your keyboard (the most important step)
Manually check against style guide
Copy-paste between platforms
Hope you didn’t miss anything
With your assistant:
Finish your article draft in
/stories/drafts/school-funding-story.md“Generate publication package for this article”
Claude creates:
Push notification (60 characters, attention-grabbing)
Twitter thread (5 tweets with key findings)
LinkedIn post (professional tone, context for policy audience)
Facebook post (community-focused angle)
Style guide check flagging any issues
All saved in
/stories/drafts/school-funding-publication-package.mdYou review, adjust, and publish
Here’s where it gets powerful: These three examples actually flow together into one system.
You start your day with the AI news digest (Example 1), which surfaces a relevant story. You use that context to prep your interview questions (Example 2). After the interview, you draft your article. Then you generate the complete publication package (Example 3).
It’s a dumbed-down example, but you get the idea. Claude remembers the full context: what was in the morning digest, what you asked in the interview, what the story is about. You’re not re-explaining at each step. You’re building on a continuous workflow.
That’s the difference between using AI as a chatbot vs. building it into your actual work system.
What it costs
Claude Pro or ChatGPT Plus subscription: $20/month
OR API access: depending on usage
Setup time: 2-4 hours for initial workspace setup
Learning curve: 1-2 weeks to proficiency
Will this save you 10 hours a month? 20? Depends entirely on your workflow and how repetitive your tasks are. Overall, the real benefit for me isn’t so much “how many hours,” but having a persistent assistant.
What This Isn’t
Before we go further, let’s be clear about what this assistant doesn’t do:
It doesn’t write your stories. You’re still the journalist. Claude Code helps with research, organization, and tedious tasks. The reporting, the narrative, the editorial judgment: that’s all you.
It doesn’t replace fact-checking. Claude Code helps structure your fact-checking workflow, but you still verify every claim with primary sources. Never cite Claude’s analysis without independent verification.
You do have to be careful about what you share. You control what Claude can access, and it only sees the folders and files you explicitly give it. That means confidential sources, sensitive documents, or unpublished investigations stay private unless you choose to expose them. We’ll go deeper into how to set this up safely in Module 5.
It doesn’t work magically. You need to set it up properly, build good workflows, and use it consistently. This is a system, not a magic wand. And you can teach it how to improve.
You’re building a system where Claude can actually remember what the hell you talked about yesterday.
That’s the unlock.
What Happens Next
If you’re reading this thinking “okay, this sounds useful but where do I even start,” that’s literally what the next module is for.
Module 2 (in the next few days) walks through the complete setup:
Installing VS Code and Claude Code (30 minutes)
Creating your journalism folder structure
Setting up your first workflow (daily monitoring)
Testing everything with a real task
Module 3 does the same for Codex, so you can compare.
By the end of Module 2, you’ll have a working system, with an actual folder structure on your computer with Claude actively helping with your work.
Then Module 4 covers the advanced stuff (MCP integrations, subagents, deep research workflows), and Module 5 tackles ethics, security, and rolling this out in your newsroom.
One More Thing
I know this sounds like a lot. But remember: you probably already have a system for managing your journalism work. Files in folders, interview transcripts, research notes, source tracking.
All we’re doing is teaching Claude where those files live and what workflows you want it to help with.
You’re not learning to code. You’re just pointing an AI assistant at the work you’re already doing and saying “help with this part. That’s it.
Questions? Drop them in the comments :) See you in Module 2 for the actual setup! 👋




Thawed some scepticism! Thanks.