Claude for marketing: which Claude does which job

Use the lightest one that does the job. Plain chat for one-off thinking, Projects when output has to sound like you, Cowork when you want files and schedules, Claude Code when a job should run the same way every time.

Here is how they compare, by marketing job. The details come from Anthropic's own docs, checked on 23 October 2026 (see Sources).

Access: chat and Projects work on a free account (Projects: up to five). Cowork needs a paid plan and the Claude Desktop app. Claude Code runs in a terminal.

If you don't code: start in Projects, add Cowork when you want recurring work. If you already built with Claude Code, skip straight there. You have the environment, and the context lives next to your product.

Write your brand context down once, in a place Claude rereads every session. That's the single biggest lever, and every surface has a slot for it.

In a Project, you add documents to the project knowledge and write project instructions. Anthropic describes both: files Claude uses as background, plus instructions that shape its responses. In Claude Code the equivalent is a CLAUDE.md file that loads at the start of every session. Cowork reads files from folders you connect, so a brand.md in the right folder does the same job.

What to put in it:

1. Who it's for, in their words. Pull real phrases from support emails, reviews and Reddit threads, not "SMB decision-makers".
2. What you sound like. Three adjectives are useless. Two paragraphs you wrote that you'd happily put your name to are not.
3. What you never say. Our list of AI tells to cut (delve, unlock, "game-changer") is a file the writing step must pass.
4. What's true. Prices, features, limits. Claude will otherwise fill the gaps with plausible fiction.

Then ask for one deliverable at a time. "Draft the launch post, 150 words, for someone who has never used a video editor" beats "plan my marketing".

Claude can read and synthesise anything you hand it: interview transcripts, a pile of reviews, ten competitor pages. It can't see search volume, your analytics or your rivals' ad spend unless something gives it that data.

Good first jobs:

• Paste 20 reviews of a competitor and ask for the repeated complaints, quoted.
• Give it your last five customer conversations and ask what words they use for the problem.
• Ask it to compare three landing pages against your positioning, line by line.

The failure mode is confident invention. Ask "how big is this market?" in plain chat and you'll get a number with no source. Say "use only what's in these files; mark anything you can't verify" and the answer gets far more useful. The same applies to competitor facts: we only publish one when it has a source and a date.

Connect it to data. Claude alone can brainstorm keywords, but it can't tell you which ones anyone searches. In Claude Code you can have it call a keyword and SERP API through scripts, then reason over the results.

This is the part we run ourselves. In this website's repo:

• scripts/keywords/discover.mjs takes a topic file, pulls Google autocomplete for every seed, gets volume and difficulty for each candidate, scores them with written rules, and checks the top ones against the live search results and Hacker News. It writes a CSV and a cluster sheet to research/keywords/. The script's own header puts volume pulls at about $0.09 per thousand keywords.
• scripts/dataforseo/serp.mjs pulls the top ten, the People Also Ask questions and the AI Overview for a query into a markdown file (about $0.004 per keyword, cached). serp-pages.mjs then fetches each ranking page and records its outline, word count and whether it has tables or code.
• Each post gets a typed "notes" file in src/content/seo/ with the competitors, a SWOT, what the post must cover, what we can add that the top five don't have, and a 16-item checklist. A page called /drafts/ renders them all.
• npm run report:daily pulls yesterday's visits from PostHog into a markdown report and a CSV row.

The research for this very guide came out of that loop. Claude pulled the search results, read the five ranking pages and told us that none of them compares the four Claudes. That gap is why this post is structured as it is.

What it hasn't done: made us rank. As of 22 October the site had roughly 265 Google impressions over 28 days, mostly for the brand name. A method that takes an afternoon is not the same as a result. Also, the loop breaks in dull ways: our keyword-data login was rejected in late October, and no model fixes an expired password.

Give it a voice file, a worked example and a banned list, then edit like a human. Generic output is what you get when the brief is generic.

What works for us:

• A voice skill. In Claude Code a skill is a SKILL.md file Claude loads when relevant, either automatically or when you type its name. We keep project skills for product facts, audience, voice and the SEO workflow in .claude/skills/. Writing a post means loading all four first.
• One real moment per piece. "It's 9pm and there are 30 unposted clips on your phone" beats a paragraph about "the importance of consistency".
• No invented proof. We never let a draft include a customer quote or stat we can't trace. Claude is willing to write one if you let it.
• A human approves. Nothing we write with an LLM ships as written. The post you're reading was drafted by Claude and approved by a person before it went live.

Disclosure matters too: if AI wrote a draft, say so where readers would expect to be told.

Reporting is where Cowork and Claude Code both earn their keep, because the task is boring, repeatable and checkable.

Cowork: point it at a folder of exports and a brief, and schedule the task. Anthropic's docs describe recurring tasks via /schedule, which run in the cloud so your computer doesn't need to be on. Anything touching local files still goes through the Claude Desktop app. Cowork is on paid plans only.

Claude Code: write the report as a script, have Claude explain odd numbers, and keep the output in the repo. Ours is the PostHog report above; each morning's file is dated, so you can see drift.

For repurposing, the useful pattern is one source, many shapes. Give Claude a blog post and ask for five LinkedIn posts, each opening on a different point, plus a short script for a video. Review each. The scripts are fine; the lines you'd never say are the ones to cut.

Yes, with some patience. You describe the job, Claude writes and runs the scripts, and you approve each command. Nobody asks you to write code.

What you do need: comfort in a terminal, a folder structure that you keep tidy, and a habit of reading what it changed. Mistakes here are less "wrong sentence" and more "overwrote a file". Use version control, and keep your brand context in a CLAUDE.md and skills so every session starts informed.

If that's too much, Projects plus a few connectors covers most of the writing and research jobs. We'd rather say that than push the terminal on everyone.

If you'd rather see the video step than read about it, you can open TopPost free and look at how an edit comes together before you decide whether it's for you.

Claude doesn't post for you, doesn't know your results, and has no video editor built in.

• Distribution. It can draft posts and, via a connector or MCP, sometimes publish. It can't create an audience.
• Judgement. It will produce ten angles. It won't tell you which one your customer cares about.
• Fresh facts. Features change quickly. Cowork, skills and connectors all changed shape during 2026, so check Anthropic's pages before relying on any limit we've stated here.
• Video. Claude can write the script, the hook and the caption. Turning footage into a finished short is a separate tool, which is the last part of this guide.

Only if you use Claude Code (or another client that takes an API key), and only with an MCP server that does the editing. That's the one place TopPost comes in, and it comes with limits.

Short answer on how it works: TopPost is a short-form video pipeline in the browser. It edits your footage, A/B tests it as trial reels, schedules the winner and turns comments into leads with auto DMs. Its MCP server lets an AI client run it. Claude can upload your footage, build and edit the video, add captions and motion graphics, send a review link, read the comments and fix them, and schedule posts. There are 45 hosted tools, and nothing to install or render on your side. Connecting looks like this:

Connect TopPost to Claude Code (create an API key first):

claude mcp add --transport http toppost https://app.toppost.ai/api/mcp --header "Authorization: Bearer <your API key>"

Any other MCP client: endpoint https://app.toppost.ai/api/mcp (streamable HTTP), header Authorization: Bearer <your API key>.

What you should know before you try it:

• Claude Code and API-key clients only. Sign-in from claude.ai or Claude Desktop (OAuth) isn't live, so Cowork and chat can't drive it. Anthropic's docs say chat connectors can take a header-based key, but we haven't confirmed TopPost that way, so we don't claim it.
• Your footage, not generated video. TopPost edits clips you record. It doesn't generate video from nothing, and it doesn't make avatars.
• Output. Vertical 9:16 MP4 only for now.
• Cost. Edits made by your own AI use no TopPost credits; you pay for storage only. Plans and storage are on the pricing page.
• How it works end to end is on How TopPost works.

An example prompt in Claude Code, once connected: "Take the three clips I uploaded this week, cut the dead air, add captions in our style, make one 9:16 short for each and put them on the calendar for Tuesday, Thursday and Saturday." You still watch each one before it goes out. If you only post a video now and then, a free editor like CapCut is probably enough (TopPost vs CapCut has the comparison).

Pick one job, not all of them.

1. Write one page of brand context: customer words, voice samples, banned phrases, true facts.
2. Put it in a Project (or a CLAUDE.md).
3. Ask for one thing you were going to put off, such as a launch post or a competitor teardown, and edit the draft.
4. If it helped, automate the boring repeat: a weekly report, a keyword pull.
5. Only then think about video.

Everything above is checkable against Anthropic's own pages, listed below with the dates we read them.