Nick Irmo
Nick Irmo
Services
Work
Insights
Resources
About
Contact
AI tools

Claude chat, Claude for work, and Claude Code: which one to use for what

I run marketing with all three, every day. They look similar and they are not. Here is the split that keeps me from using a chat window for a job that needs an agent, or an agent for a job that needs a conversation.

October 26, 20268 min readBy Nick Irmo
In shortI run marketing with all three, every day. They look similar and they are not. Here is the split that keeps me from using a chat window for a job that needs an agent, or an agent for a job that needs a conversation.

Three products, one name, and a lot of marketers using the wrong one for the job. I have spent most of 2026 running a two-firm marketing function with Claude in all its forms, and the thing that made it useful was not prompting. It was learning which door to walk through.

A note before the split: Anthropic has folded what it used to call Cowork into the main Claude app, so if you still see the old name in a tutorial, it is describing the same work mode I use below. The capabilities are the same; the label moved.

Claude chat: the thinking partner

The chat window is a conversation. It is the right tool when the output is a decision, a draft, or an understanding, and when you expect to go back and forth.

Use it for: positioning arguments, rewriting a paragraph until it sounds like you, pressure-testing a plan, explaining a regulation in plain language, turning a messy brief into a clean one, and asking what you are not seeing. Give it context in the first message and treat the next ten as editing.

Do not use it for: anything that touches more than one file, anything you want to come back to next week, or any job where "done" means a thing exists somewhere other than the chat.

Claude for work: the one that finishes things

This is the mode where Claude has your files, your connected apps and a long leash, and the output is a deliverable rather than a reply. A deck. A spreadsheet with formulas. A research report with sources. A website page. A batch of posts scheduled in Buffer. A folder on your Mac reorganized. The job runs for minutes, sometimes longer, and you get the result, not a running commentary.

Use it for: anything with a file at the end. Weekly reporting pulled from GA4 and the CRM. A content package: article, LinkedIn post, carousel, images. Campaign research across twenty competitor sites. Cleaning a client folder. Building a dashboard. Rewriting fifty product descriptions with a style guide attached. My weekly thought-leadership pack, including this article, is produced this way: research, draft, assets, published to the site, in one run.

The skill that matters here is the brief, not the prompt. Say what done looks like, where the inputs are, what to leave alone, and what you want to decide yourself. A good brief to the work mode reads like a good brief to a contractor, because that is what it is.

Do not use it for: a single question, or anything you have not thought through enough to describe the finished product. It will build what you asked for, quickly, which is only useful if you asked for the right thing.

Claude Code: the one for anything that runs

Claude Code lives in the terminal, the IDE, or a browser session attached to a repository, and it changes code, runs it, tests it and ships it. It is a developer tool by design, and it is also the best tool I have for marketing work that has code underneath it, which is more of it than people admit.

Use it for: website builds and fixes, tracking and tag implementation, HubSpot or Salesforce automation that needs scripting, data pipelines, landing page generators, static sites, schema markup, performance work, the Looker Studio connector nobody wants to configure by hand. It reads the whole codebase before it acts, which is the difference between an edit and a guess.

Do not use it for: copy, strategy, or anything a non-technical colleague needs to open and edit afterward. It will produce them, and they will live in a repository where nobody on the marketing team will ever look.

The rule of thumb

Ask what shape the output is.

  • A thought, a draft, a decision: chat.
  • A file, a deck, a report, a change inside an app you use: the work mode.
  • Something that compiles, deploys, or runs on a schedule: Code.

If you are not sure, start in chat, get to the point where you can describe the finished thing in three sentences, then hand that description to the work mode or to Code. Most wasted hours I have seen come from skipping that step: a vague request handed to an agent, or a twelve-step deliverable attempted one message at a time in a chat window.

What this changes for a small team

A one- or two-person marketing department used to be limited by hands. With the three tools split correctly, it is limited by judgment: knowing what to ask for and whether what came back is right. That is a better constraint to have, and it is the one the IRMO Playbook is built around. Identify and Roadmap are chat work. Market is work-mode and Code. Optimize is all three, every week.

The tools will keep changing names. The split will not.

Key takeaway

Chat is for thinking. The work mode is for finishing things that live in files and apps. Code is for anything that compiles or deploys. Pick by the shape of the output, not by how clever the request is.

Next step

Want this thinking applied to your pipeline?

Send a note No pitch deck. A reply within one business day.
Want this applied to your pipeline?