SHORT ANSWER
Using Claude Code instead of the Claude web app doesn't make the model more factually reliable. Claude Sonnet is Claude Sonnet either way. What the CLI gives you isn't accuracy — it's file access, tool use, and the ability to chain steps together. Those matter, but they don't solve the hallucination problem. Fact-checking and editorial judgment are still your job, the same way they are in the browser.
Why the CLI doesn't change the underlying problem
An LLM hallucinates when it produces something that sounds plausible but isn't actually true: a wrong quote attribution, a source that doesn't exist, a date off by six months, a statute that was never passed. That happens inside the model, before any file gets written. The terminal doesn't sit between the model and your disk policing for factual errors — it just hands you what the model produced and saves it wherever you said to save it.
Running the same prompt in a terminal instead of a browser doesn't change the model's reasoning, its training data, or the patterns that produce hallucinations. If you asked Claude on the web to "summarize this council meeting" and it made up a vote count, running the same prompt in Claude Code will make up a vote count too.
The CLI does give you better tools for verification — real files you can diff, source documents you can reread, prompts you can reuse. But "better tools for verification" and "fewer hallucinations" are not the same thing.
What the course actually teaches about verification
Module 4 introduces a pattern Joe calls cross-model code review: you call a second CLI tool — Copilot CLI, Codex CLI, or Gemini CLI — from inside your main Claude Code session using a -p flag. The second tool runs in its own isolated context window and gives you an independent second opinion on whatever Claude just built.
In the Module 4 video, Joe demonstrates this by asking Claude to review a branch of code, then calling Copilot CLI and then Codex CLI to review the same branch as second and third passes. Codex catches two high-severity bugs that both Claude and Copilot had missed. The framing is "different models catch different bugs" — and the pattern works because the second model re-reads the underlying code with fresh eyes, not Claude's summary of what it did.
That's a genuine accuracy boost — but it's specifically about code review. It's the pattern you'd use to catch a bug in a pipeline script before you commit it, not the pattern you'd use to fact-check a news story.
The hard line: editorial judgment stays with you
This course is about using CLI tools to build the scripts, skills, pipelines, and context files that shape your workflow. It is not about delegating the reporting, writing, or editing of your journalism to an AI model. Module 4 makes this explicit: the journalist-as-manager shift is about managing the code side of your workflow the way a non-developer works with a developer. Reporting, writing, editing, and verification of factual claims stay with the journalist.
That's not a disclaimer written to cover the Knight Center's legal department. It's the honest answer to what these tools can and can't do right now. A second model that's also trained on the same internet is not an independent source. Two LLMs agreeing that a quote is real doesn't make the quote real. The only thing that makes a fact true is the source it came from, and the only thing that verifies a fact is a human who reads the source and confirms it.
The journalist in front of the keyboard is the ground truth. The CLI is a faster way to organize your notes.
What actually helps, in practice
- ›Save source documents to disk. One of the biggest advantages of the CLI over the web interface is that when Claude reads a web page or a PDF, you can keep the extracted text. Read the source file, not just Claude's summary.
- ›Ask for inline source links for every claim. Same pattern Eva Wolfangel used in the forums on her research prompt: after Claude writes the first draft, ask it to add a source link next to every factual claim. Then open each link and read it. If the link doesn't exist, the claim is suspect. If the link exists but doesn't say what Claude said it said, the claim is suspect.
- ›Keep your CLAUDE.md specific about sources. Generic instructions like "verify facts" don't change output. Specific ones like "flag any dollar amount or vote count that isn't directly attributable to a document in the
sources/folder" do. - ›For code, use Module 4's cross-model review pattern. If Claude wrote a script that filters your dataset, ask Copilot or Codex to review that script before you trust its output. Codex found real bugs in Joe's demo that Claude and Copilot both missed.
THE THING NOBODY LIKES TO SAY OUT LOUD
There is no tool, flag, plugin, or subscription that turns an LLM into a reliable fact checker. The models are useful for a lot of things — reading long documents, extracting structured data, organizing notes, building pipelines — but they are not trustworthy as standalone verifiers of factual claims. Treat their output the way you'd treat a fast, confident intern who hasn't been on the beat long enough to know when they're wrong.
GO DEEPER: The cross-model review pattern is covered in Module 4. The broader framing — what these tools can and can't do for journalism — runs through the whole course but lands hardest in Module 4's framing videos.