Pick Claude Cowork or Claude Code by what the work has to leave behind, not by whether you write code for a living. If the job ends as a deck, spreadsheet, or research brief, Cowork is the right surface. If the job has to survive as a reusable system (scripts, skills, hooks, tests, git history, CRM write policy), GTM operators should use Claude Code.
That split still holds after Anthropic started folding Cowork into chat. On September 16, 2026, Claude Cowork and chat became one Claude for Pro and Max plans. You ask for a weekly pipeline recap and five slides in the same conversation. Claude decides whether that is a quick answer or a longer task. Team and Enterprise orgs still keep chat and Cowork separate for now. Claude Code did not disappear into that merge. It is still a different operating surface: a repo, a permission model, and work that can be reviewed like software.
What is the difference between Claude Cowork and Claude Code?
Anthropic built both on the same agent loop. Cowork applies that loop to files, folders, and connected apps. Claude Code applies it to a codebase.
Anthropic's own tutorial is blunt about Cowork. Chat is the surface where you stay in every turn and assemble the output yourself. Cowork is the surface where you describe a deliverable, Claude works across files and tools, and you come back to a finished file or an action taken. The same page says Cowork runs the Claude Code engine on your files and apps instead of a repo.
Claude Code's current docs put the leftover on the other side. It edits files, runs commands, works with git, opens pull requests, and can run in CI. You add CLAUDE.md for always-on project rules, skills for repeatable workflows, hooks for events that must fire every time, and MCP for external systems. That is why a HubSpot property cleanup in a git repo behaves differently from the same cleanup as a one-off spreadsheet in a desktop folder.
Cowork also does not inherit your Claude Code CLI setup by accident. Cowork loads connectors, skills, and plugins from your claude.ai account. It does not read the CLI ~/.claude directory on your machine. If the GTM skill only lives in the terminal profile, Cowork will not see it until you add it in Customize. Operators miss that and then blame the model.
When should GTM operators use Claude Cowork?
Use Cowork when the output is a file a human will open, not a system another operator has to rerun.
Good Cowork jobs:
- Turn last week's call notes and Slack threads into a one-page pipeline recap.
- Build a QBR deck from a folder of exports.
- Clean a local CSV before anyone imports it.
- Research five accounts and drop a brief next to the source PDFs.
- Recurring Monday status that should land as a doc, not as a production workflow.
Anthropic's Cowork architecture still matters here. Cloud Cowork sessions run in an isolated sandbox on Anthropic's servers. Local files only come in through the desktop app, for folders the person connected, while that app is online. Connector tokens stay server-side. That is a better design than dumping CRM credentials into a chat paste. It is still not the same thing as a reviewed pull request in your operating repo.
Cowork is also the wrong default for anything you will need to defend in six weeks. The deck looks done. The logic that produced the numbers lives in the conversation. Next month's intern cannot rerun it.
When should GTM operators use Claude Code?
Use Claude Code when the leftover has to be the team's system.
Good Claude Code jobs:
- A HubSpot property and workflow audit that produces a markdown report plus the scripts used to pull the export.
- A lead-routing rule set stored as files, with tests for the edge cases sales already hit.
- A transcript-to-task skill that the next operator can run without rebuilding the prompt.
- A staged CRM write path: read-only first, draft packet second, write only after a human approval.
- Recurring pipeline hygiene that should fail in CI if the definitions drift.
That is the same argument as how to use Claude Code for GTM operations. The interesting part is not that it can write Python. The interesting part is that the work sits next to context files, instruction files, and outputs that survive the session. If you need the model to remember the operating rules, put them in the repo. Do not hope the chat remembers.
Claude Code is also where hooks live. A skill can ask the model to follow a CRM write policy. A hook can block the write when the policy is ignored. GTM teams that skip that layer end up with a helpful agent that still overwrites owner, lifecycle, or amount because nobody put a hard stop in front of the tool.
What are GTM teams using each surface for?
In March and April 2026, Kyle Poyar and Maja Voje surveyed 200 GTM operators who had used Claude Code or Cowork. It is an early-adopter crowd, and they say so. The split is still useful. Primary Claude product was Chat for 30%, Code for 31%, and Cowork for 32%. 92% said the products saved time. 67% said they enabled something that had been impossible before.
The impact split matches the leftover test. Cowork users ranked content creation first (33%), including social posts, sites, and structured content pipelines. Code users ranked GTM engines and prospecting first (28%), including SDR tools, segment messaging, and signal-based outbound sequences.
"Think in Chat, build in Code, and operate in Cowork."
Maja Voje, GTM Strategist, as quoted in Kyle Poyar's 2026 Claude for GTM Pulse Report
That shorthand is fine as a starting map. It breaks if "operate" means "this is now how lead routing works for the team." Operating a personal Monday recap can live in Cowork. Operating the company's routing rules belongs in Claude Code, in a repo other people can clone.
The same survey is honest about the tax. Credits and usage limits showed up for roughly half of respondents. Sharing is worse. One operator said a useful skill still has to be rebuilt by each coworker. Another said you have to hope colleagues update the skill after you change it. That is the GTM version of shadow IT. If the workflow matters, put it in git.
| Job | Better surface | Leftover you should keep |
|---|---|---|
| Weekly pipeline recap or QBR slides | Cowork / chat | The finished doc, plus the folder of source exports |
| One-off account research brief | Cowork / chat | The brief next to the source files |
| CRM field and workflow audit | Claude Code | Report, scripts, and the instruction file that defined "good" |
| Lead routing or SLA clocks | Claude Code | Rules, tests, exception path, and a reviewed change |
| Recurring content drafts with a voice guide | Cowork first, Code if it becomes a pipeline | Voice rules in a file the next run can load |
| Agent with CRM write access | Claude Code, staged | Permission map, approval packet, audit trail |
Why did Anthropic merge Cowork and chat?
Because people were wasting time picking a mode.
The September 16 product note says users got stuck deciding where a task belonged, and work started in one place did not carry into the other. The merge is a UI fix for that. It is not a claim that Claude Code is now optional.
If anything, the merge makes the Code decision cleaner. Chat and Cowork are collapsing into "ask Claude, get a file or an answer." Claude Code stays the place where GTM work becomes software: CLAUDE.md, skills, hooks, subagents, MCP, git. Anthropic's feature map still tells you to start with CLAUDE.md for conventions, then add a skill when you keep pasting the same playbook, then add a hook when something must happen every time without asking.
"Claude Code should always be a WIP. Optimizing your AI environment is ongoing work, not a one-time setup."
Sebastian Silva, Founder, HigherOps
A merge in the chat product does not finish that WIP. It removes a mode picker. The operating repo is still the memory layer.
How should a RevOps team split the work?
Start with the artifact, then pick the surface.
If the next useful thing is a file for a meeting, use Cowork or the merged chat. Keep the source folder. Do not connect production HubSpot write tools to that session.
If the next useful thing is a rule the team will rerun, open Claude Code in the GTM repo. Put company context and write policy in context files. Put the procedure in a skill. Put the "never write these fields" rule in a hook or a human approval step. Keep an audit trail of what the agent proposed and what a human accepted.
A practical first week:
- Pick one leftover that already repeats. Pipeline recap, property audit, or routing QA. Not "use AI more."
- If it is a recap, run it in Cowork against a connected folder. Save the output next to the inputs.
- If it is an audit or a rule, run it in Claude Code against exported HubSpot data and a CLAUDE.md that states the definitions.
- Do not give either surface a write token on day one. Draft a packet. A human clicks.
- If two people now need the same workflow, move it into the repo. Stop shipping skills as Slack screenshots.
The failure mode I keep seeing is a RevOps lead who lives in Cowork because the terminal feels like engineering theater. They get fast decks. Six weeks later nobody can explain why MQL means three different things, because the scoring logic never left the chat. Cowork did its job. The team used it for the job Claude Code was for.
The other failure is the opposite. Someone opens Claude Code, skips project setup, and treats the CLI like a smarter ChatGPT. No CLAUDE.md. No skill. No review. Then they paste a private app token into the session and call it automation. That is not Claude Code for GTM. That is a faster way to make a mess.
Frequently asked questions
Do GTM operators need to know how to code to use Claude Code?
No. They need to be willing to work in a folder with files, diffs, and review. If you can maintain a HubSpot workflow, a reporting definition, and a naming convention, you can maintain a GTM repo. The coding is often the least interesting part. The operating rules are the work.
Should we shut down Cowork now that chat can do the same jobs?
No. On Team and Enterprise, Cowork is still its own surface. On Pro and Max, the merge means you stop picking a tab. Use that surface for finished files. Keep Claude Code for systems. The product names will keep moving. The leftover test will not.
Can Cowork replace Claude Code for CRM operations?
Not if the CRM work has to be rerun, reviewed, or handed to another operator. Cowork can draft the analysis from a local export. Claude Code is where that analysis becomes a checked-in procedure. If you need both, run Cowork for the first artifact, then move the stable parts into the repo.
What should never go into either surface?
Production write access on owner, lifecycle, amount, close date, or customer status, until there is a permission map and a human approval step. Also skip pasting private app tokens into a chat. If the agent needs HubSpot, use a scoped connection and log what it changed.
Key takeaways
- Pick Cowork or Claude Code by leftover work, not by job title.
- Cowork (and merged chat) is for finished files. Claude Code is for reusable GTM systems.
- Anthropic's September 2026 chat and Cowork merge removes a mode picker. It does not retire Claude Code.
- In a 2026 survey of 200 GTM operators, Cowork's biggest impact was content. Code's biggest impact was GTM engines.
- Skills that only live in one person's chat are shadow IT. Put team workflows in git.
- Do not give either surface unreviewed CRM write access.