Most B2B marketing teams are stuck in a writing bottleneck they do not even see as a bottleneck anymore. A single 1,000-word blog post takes about 4.2 hours to research, write, and edit, per Affinco's April 2026 content statistics, which do not name the survey behind the figure. Multiply that by two or three posts a week, and you have one person spending half their working hours on content production. That is before you factor in the distribution: the LinkedIn posts, the email version, the social snippets. The maths says most teams are publishing less than they should because the production pipeline simply cannot keep up with the strategy.
The numbers back this up. 85 percent of marketers now use AI for content creation, up from 61 percent in 2023, according to the same Affinco roundup, and 94 percent plan to use it for content creation, blog posts included. Marketing teams using AI report 44 percent higher productivity and save an average of 11 hours per week, a HubSpot figure quoted in DataRefs' July 2026 AI marketing statistics. Adoption is near universal now. What separates results is whether a team uses AI as a one-shot drafting tool or runs an actual agent across the whole pipeline.
This post walks through building that agent: a persistent Claude Cowork setup with three skill files that together form a research, write, and edit pipeline. You configure it once, and it follows the same rules on every post. After that, you give it a topic and get back a researched draft in your brand voice, with the editing skill's flags attached for a human to check. Anyone with Claude Pro and a free afternoon can build this.
An AI blog writing agent removes the three hours of research and first-draft writing that happen before the human editor ever touches the content. The human still makes the final decisions. The agent just gets them to the decision point faster.
If you have already set up content repurposing with Claude Cowork and Buffer MCP, the pattern here will feel familiar. The same skill-based approach that turns one blog post into ten social pieces is what turns a topic into a researched, drafted, and edited blog post. And the ten-layer agent stack principles apply here too: the brain reasons, the skills define the identity, and the tools do the execution.
The pipeline needs one paid account and a text editor
You need three things. A Claude Pro account, which costs USD 20 a month billed monthly (USD 17 a month on the annual plan) and includes Cowork, per Anthropic's Cowork help page. A text editor to write skill files, anything from VS Code to Notepad works. And optionally, Buffer if you want the agent to queue the social posts that announce each article. Buffer's free plan connects up to three channels. That is the full list. No API keys, no custom code, no separate hosting.
Cowork is the part of Claude that carries out multi-step tasks on its own: it reads and writes files, uses connectors such as Buffer, and loads skills, which are packaged instruction files that Claude picks up whenever a request matches their description. That last point is what makes this an agent you set up once. Every blog post it writes follows the same rules without you pasting them in again.
Anthropic changed how Cowork works while this guide was live, so check these three points before you build. We read each of the help pages below on 4 October 2026.
| Part | Where it lives | Anthropic's page |
|---|---|---|
| Research, voice and editing rules | Three skills, each a folder with a SKILL.md file, zipped and uploaded under Customize, then Skills |
How to create custom skills |
| Standing instructions and memory | A Cowork project, which keeps instructions and remembers earlier tasks in the same project | Projects in Cowork |
| The weekly run | A scheduled task, which runs in the cloud and can use connectors and files saved to your Claude account, but cannot be tied to a folder on your computer | Schedule recurring tasks |
Two more changes matter here. From 6 October 2026, new Cowork tasks run in the cloud and the "Only on your computer" setting is removed. And Anthropic is merging Cowork and chat into one Claude, starting with Pro and Max, so the separate Cowork tab may already be gone from your app. Your projects, skills, connectors and scheduled tasks carry over, and the steps below work either way.
Step 1: Set up a Claude Cowork project for blog writing
Create the project and a folder for your skills
In the Claude app, find Projects in the left sidebar, click the + button and choose to start from scratch. Call the project blog-agent. In its instructions, write one line saying what the project is for, such as "Research, write and edit posts for our company blog". Choose start from scratch over linking a folder on your computer, because a project tied to a local folder stays on that computer and the weekly run in Step 6 cannot use it.
On your own machine, make a working folder such as ~/Documents/blog-agent-skills/. You will write the three skill files there, one sub-folder per skill, and zip each one to upload it.
If you want the agent to queue social posts, connect Buffer now. Open Customize, then Connectors, search for Buffer and click Connect, then sign in to Buffer and approve access. Buffer documents this on its Claude integration guide. If Buffer is not in your connector list, choose Add custom connector and paste https://mcp.buffer.com/mcp, Buffer's own MCP server address. The connector lets Claude see your scheduled posts and draft new ones for your channels.
Verify it works: ask Claude, inside the project, "Which Buffer channels can you see?" If it lists your channels, the connector is live. If it says it has no Buffer tools, the connector is not switched on for this task.
Step 2: Build a research skill that searches before writing
Create the research skill file
Inside your working folder, create a sub-folder called blog-research and a file inside it called SKILL.md. This skill tells Claude how to research before writing a single word. The two lines between the dashes at the top are required: Claude reads the description to decide when to use the skill, so write it the way you would ask for the task. Here is a starting template:
--- name: blog-research description: Research a blog topic before drafting. Use when asked to research or write a blog post. --- # Research Skill ## Purpose Before writing any blog post, gather evidence from the web. ## Research rules 1. Search for three statistics from named sources published in 2025 or 2026. 2. Find two competitor articles on the same topic and note what they miss. 3. Identify one unique angle not covered by competitors. 4. Collect the source URL, publication date, and exact wording for every claim. ## Output format Return your findings in this structure: - Primary stat (source, date, exact quote) - Secondary stats (source, date, exact quote) - Competitor gap: what the top results miss - Unique angle: what this post adds that others do not
Zip the blog-research folder so the folder itself sits at the root of the zip, then upload it under Customize, Skills, the + button, Create skill, Upload a skill. Once it is enabled, Claude loads it whenever a request matches the description, so you do not need to paste the rules again. The output format matters more than it looks, because the editing skill in Step 4 checks every statistic against the source, date and exact wording this step records.
Step 3: Build a writing skill with your brand voice
Write the brand voice and structure rules
This is the skill that matters most. Without it, Claude writes generic blog posts. With it, every post follows the same tone, structure, and rules. Create brand-voice/SKILL.md in your working folder:
---
name: brand-voice
description: Our blog's tone, audience and structure. Use when drafting any blog post.
---
# Brand Voice Skill
## Tone
Professional but direct. No jargon. Company voice ("we"), not personal ("I").
Keep paragraphs to 3 to 5 sentences. Vary sentence length.
## Audience
B2B marketing leaders at companies with 10 to 200 employees.
They know marketing but are evaluating AI tools. Speak to their outcomes, not the technology.
## Structure
Every blog post follows this structure:
1. Opening paragraph: arresting stat or scenario, then the solution promise
2. "What you need" section: tools and cost, before any setup steps
3. Numbered steps with exact paths, commands, and verification checks
4. Results section: time saved, output comparison
5. FAQ section with 5 to 6 questions
## Banned phrases
"synergy", "drill down", "circle back", "touch base", "value-add",
"move the needle", "in today's fast-paced", "it is not X, it is Y"
## Capitalisation and formatting
Normal capitalisation rules. Proper nouns capitalised. Sentence case headings.
No em dashes. Use commas and full stops instead.
The rules above are a starting point. The part that makes output sound like you is examples: paste two or three paragraphs from posts you are proud of under a heading such as "Sounds like us", and one paragraph you would never publish under "Does not sound like us". We could not find a published test that measures how much a voice file improves output, so judge it on your own drafts and edit the file whenever a draft misses. The file is plain markdown. If you can write a Notion page, you can write a Claude skill. Zip and upload it the same way as the research skill.
Step 4: Build an editing skill that checks facts and tone
Create the editing and quality-control skill
The editing step is where most AI content pipelines fall apart. The draft looks good on first read, but the stats are unsourced, the tone drifts, and a few sentences sound unmistakably AI-generated. An editing skill catches this before a human ever looks at it. Create blog-edit/SKILL.md, then zip and upload it like the other two:
--- name: blog-edit description: Check a blog draft for unsourced facts, tone and AI tells. Use after drafting a blog post. --- # Editing Skill ## Purpose Review every blog draft before it reaches a human editor. ## Checks (run in order) 1. Fact check: every statistic must have a named source. Flag any stat without one. 2. TOV check: scan for banned phrases from the brand voice skill. Flag every violation. 3. AI-tell check: flag sentences that start with "It is not X, it is Y", colon constructions introducing clauses, and consecutive sentences with identical structure. 4. Link check: every external source name should be a hyperlink. Flag any that are not. 5. Readability: flag any sentence over 30 words. ## Action For each flagged issue, rewrite the sentence and present both versions. The human editor picks which version to keep.
Running this editing step as a gate before the human review pass changes what the human is actually doing. Instead of proofreading every sentence, they are checking the handful of flagged issues and making the calls the agent cannot: does this angle actually serve the reader, does this claim need a stronger source, does this sound like us. That is a strategic review, not a copy edit, and it takes less time precisely because the mechanical pass is already done.
Step 5: Connect the pipeline and write your first post
Run the full research, write, and edit pipeline
Now the three skills are in place. Start a task inside the blog-agent project and give it this prompt:
"Research and write a blog post about [your topic]. Use the blog-research skill, then the brand-voice skill, then the blog-edit skill. Research first, write second, edit third. Return the final draft with the editing annotations."
Naming the skills in the prompt is a belt-and-braces step, since Claude would usually pick them up from their descriptions anyway. It searches for statistics and competitor articles, writes the post to your voice rules, then runs the editing checks against its own draft and lists what it flagged. How long that takes depends mostly on how much web research the first skill asks for, so time your first few runs before you promise anyone a turnaround. What you get back has been researched, written to your rules and checked once, and a human still reviews it before it goes live.
Buffer publishes to social channels, so it does not post the article to your website. What it can do is queue the announcement. Add a fourth instruction such as "Draft a LinkedIn post announcing this article and schedule it in Buffer for next Tuesday at 9am", and Claude creates the post in your Buffer queue through the connector. Publishing the article itself still goes through your CMS.
Step 6: Automate with weekly scheduling
Set up a recurring content cadence
Once the pipeline works for one post, make it recurring. Scheduled tasks run in the cloud, which means they keep running when your laptop is asleep, but they cannot read or write a folder on your computer. Keep the topic list and the drafts somewhere the task can reach, such as a Google Drive or Notion connector, or files saved to your Claude account. Then, in a task inside the project, type:
/schedule Every Monday morning, open the "Blog topics" doc in Google Drive. Take the first topic not marked done, run the blog-research, brand-voice and blog-edit skills on it, save the draft as a new doc in the "Blog drafts" folder, and mark the topic done.
Claude asks you to confirm the details. Anthropic's scheduling page lists hourly, daily, weekly, weekday and manual runs, available on every paid plan, and you can review or pause the task from Scheduled in the sidebar. By Monday afternoon a researched and checked draft is waiting in the drafts folder, and your job is the review: does the angle serve the reader, does each claim hold up, does it sound like you.
For teams already running content repurposing with the Claude Cowork and Buffer MCP workflow, this blog writing agent slots in before the repurposing step. The agent writes the blog post. The repurposing pipeline turns it into ten social pieces, and the human time in the whole chain sits in two review passes, one on the article and one on the social posts.
What the time saving looks like, and the part to measure yourself
Affinco's table puts a 1,000-word post at 4.2 hours done by hand and 1.8 hours with AI assistance. That second figure is for a person drafting with AI, before any skills or scheduling. We have not found a published measurement for a fully skilled pipeline like this one, and we would rather not invent one. The number worth tracking is your own review time per draft. If review is still taking more than an hour, the brand voice skill usually needs more examples, and the editing skill needs a check for whatever you keep fixing by hand.
Marketing teams using AI report publishing 42 percent more content monthly, a median of 17 articles versus 12 without AI, according to Affinco's AI content creation data. And 84 percent of marketers say AI improved the speed of content delivery, per CoSchedule data cited by Arvow's 2026 statistics roundup. Neither source separates teams running a full pipeline from teams drafting with a chat window, so read both as a floor.
| Tool | Cost | Required? |
|---|---|---|
| Claude Pro | USD 20/month, or USD 17/month billed annually | Yes |
| Buffer | Free (up to 3 channels) | Only for queuing social posts |
| n8n Starter (cloud) | EUR 20/month billed annually (self-hosted Community edition is free) | Only for extra routing |
Prices are from Claude's pricing page, Buffer's pricing page and n8n's pricing page, read on 4 October 2026. The minimum setup is Claude Pro alone. How many posts it supports a week depends on your Pro usage limits and on how long your review takes, and Anthropic does not publish a post count.
Frequently asked questions
How much does it cost to build a blog writing agent with Claude?
Claude Pro costs USD 20 a month billed monthly, or USD 17 a month on the annual plan. Buffer is free for up to three channels. n8n is optional, and its Starter cloud plan is EUR 20 a month billed annually. If you run the writing through the Claude API instead, Claude's pricing page lists Sonnet 5.5 at USD 10 per million output tokens, so the roughly 2,700 tokens of a 2,000-word draft cost about three US cents. Web research adds input tokens on top of that, and we have not metered a full run.
Can I use ChatGPT instead of Claude for automated blog writing?
Yes, in principle. OpenAI launched ChatGPT Work in July 2026 as its own task agent, and the research, write and edit split in this guide is plain instructions that you can move across. We have not built this pipeline in ChatGPT Work, so we cannot say how its setup for reusable instructions and scheduling compares.
How much time does an AI blog writing agent actually save?
Affinco puts a 1,000-word post at 4.2 hours by hand and 1.8 hours with AI assistance, and separately says marketers save roughly 3 hours per piece. Affinco does not name the surveys behind either figure. We have not found a published measurement for a fully skilled pipeline, so track your own review time per draft.
Do I need to write code to set this up?
No. Each skill is a plain-English markdown file in its own folder, which you zip and upload under Customize, then Skills. The Buffer connection is a URL you paste into the Connectors settings, followed by a Buffer sign-in. There is no JavaScript, Python or API key involved.
How many blog posts can this agent produce per week?
Anthropic does not publish a post count for Claude Pro, and the limit you hit first is usually your own review time. Start with one scheduled post a week, time the review, and add a second run only when the first draft needs little more than fact-checking.
What if Claude writes content that sounds generic?
Generic output usually means the brand voice skill has rules but no examples. Add two or three paragraphs from posts you like and one you would never publish, then rerun the same topic and compare. Edit the skill file each time a draft misses, so the fix carries into every later post.