Using Agents to Build an AI Assisted Content Pipeline

Do you know someone in business who is “good at social”? Around Fort William we have Anja, who runs Great Glen Charcuterie. The local business community speak with reverence and awe about her ability to regularly and consistently post a stream of top quality posts to her Instagram page (not to mention that the charcuterie is amazing too!).

I’m not one of those people!

Quite honestly, I struggle with Social Media. I know it’s important but there always seems like something more useful to do. Clients take priority, of course, but even when I’m between gigs or there is nothing urgent to do for a client, I still find ways to put off writing blog posts and posting to LinkedIn.

But surely this is page 1 of the AI playbook? Is generating blog and social posts not bread and butter for these fancy AI models?

Well, yes and no. Anyone can ask Claude to pump out a blog article and a couple of social posts about the latest hot topic but, as we all know, lots of people do exactly that. You’ve seen the posts and you can spot it a mile away. Overuse of emojis, unnecessary drama, em-dashes all over the place. It’s a sure sign that someone has run off a quick post using AI and slapped it up on their website or a social feed. Apart from the fact many people find it off-putting, it’s really hard to stand out from the crowd when you sound like everyone else in the crowd.

Put simply, if you want to stand out from AI slop, don’t post AI slop.

On the other hand, you can’t ignore the fact that AI is a massive time saver and that it does have the potential to create good content - or at least help with a lot of the grunt work of writing a post. And it’s getting harder to justify spending days on one piece of content when AI can fast-track it.

There has to be a middle ground.

That’s exactly how I feel, and I think I’ve finally found something that works for me. I made an AI assisted pipeline using Claude and some popular online apps. I’m going to explain it here using a real post as the worked example: Making Tax Digital.

First, what do I want?

My content pipeline is to post an article to my blog and then use it to generate social media posts. My blog posts are often research based and can be time consuming to get right. Social posts are easier

I want a system that:

• Is controlled using AI Agents - this is the secret to automation
• Is repeatable. I can’t improve it if it’s different every time.
• Uses AI to assist me, not take over. I need to be able to step in and change anything about a post whenever I need to.
• Makes managing the voice used in AI writing really easy
• Schedules social posts in advance so I know I have something posted when I’m busy

I’m deferring, for now, the following:

• Measuring the impact. I can do some ad hoc observations (“oh, that post did well” or “I feel like it’s getting easier”) but capturing and tracking metrics can come later. After all, I can’t measure something that doesn’t exist or fails all the time.
• Full automation. I’m happy to do some things manually as long as I’m getting content out more easily and regularly and I’m slowly automating more of the process

So what’s an agent?

Having Agents like Claude Cowork to drive the process is key to automation. But what are agents, anyway?

If you’ve connected ChatGPT or Claude to your Gmail or your Google Drive, you’ve already seen the beginning of this. The chat can read your inbox, open a spreadsheet, pull out what you asked for. Useful, but you’re still the one driving: you ask, it answers, you copy the answer somewhere else, you ask the next thing.

An agent takes a goal and pursues it. It works in a loop: choose an action, run it, read the result, decide what to do next. A chatbot answers the message you sent and waits for the next one. That’s the whole difference.

In practice, that means a handful of things a chat window on its own can’t do:

• It orchestrates. Instead of you copying between systems, the AI moves the information itself. Pull this record, format it, put it there, update the status.
• It reads your files. Not just what you paste in - actual documents on your computer or in your cloud storage.
• It joins apps together. Your database, your scheduling tool, your website. One instruction can touch all three.
• It works on a schedule. A daily email triage that creates follow-up items in your to-do list, running at 7am whether or not you’re at your desk. The work happens on the provider’s machines, so your laptop can be switched off.
• It remembers how you like things done. Most agent platforms let you save a set of instructions as a reusable “skill”. A skill is nothing more exotic than a prompt you’ve written down properly so you can use it again, or hand it to someone else. “Here’s how I triage email” becomes a thing the agent can do on request, the same way every time.

The line between chat and agent blurs with every platform release. The practical difference is who does the legwork: with a chat, you do it and the AI helps. With an agent, the AI does it and you decide.

What I really wanted

Here’s the plan.

One to three properly researched blog posts a week, depending on how much client work is on. Assisted by AI, not written by AI. In practice that means iterative drafting against a very well-defined description of how I write, so that when the AI edits a paragraph it doesn’t come back sounding like a press release. I still read every word and tune the final output. The Making Tax Digital post is one that came out of this process.

A cover image for each post, generated with Nano Banana 2 in a visual style that matches my brand. Sometimes diagrams for the body too.

Then, once the post and images are finished, they go into a central database, and three agents take it from there.

The Social Media Manager

This one reads the finished post and drafts candidate LinkedIn posts for me to review. Each one comes with a post body drawn from the article, a caption for the image, and hashtags, all written against the same voice description the blog drafting uses.

LinkedIn is the only platform I’m targeting right now. That’s deliberate. It’s where the business community really is, and I’d rather get one channel right than do four badly. Other platforms can come once the process is proven.

The agent then takes the cover image and the caption and generates a thumbnail for each post.

I read the drafts, tweak what needs tweaking, and the agent adds them to an automated posting schedule.

The Social Media Recycler

I want something going out every day, regardless of whether I’ve published anything new.

So a scheduled task runs once a week, looks at the next fortnight’s posting schedule, finds the days with nothing on them, and fills the gaps from a pool of general-purpose posts I’ve written and approved in advance. No new content invented, no gaps left.

The Blogging Platform Agent

The last one pulls the post and images out of the database, formats them into my blog’s house style, writes a TL;DR section for the top (again, in my voice), and publishes.

This is formatting drudgery: fiddly, repetitive, and easy to get subtly wrong. An agent gets it right every time.

How it’s put together

Almost none of this is exotic.

A central database. Mine’s in Airtable, holding articles, voice definitions, and assets like the social thumbnails. It’s all text. Nothing clever.

A scheduling tool for social posts. I use Buffer. Plenty of alternatives do the same job.

An agent platform that connects to both. I’m using Claude Cowork with its Airtable and Buffer connectors, which is what ties the scheduling and posting together. Cowork is built for people who need to get things done, not for developers.

A blogging platform. I run my own, so I wrote a custom skill to post to it. If you’re on WordPress or one of the mainstream platforms, there’s likely a connector already built and you can skip that entirely.

Something to write in. Gemini, Word with Copilot, whatever you already use. I built myself a small writing app, but that’s an advanced detour, not a requirement. The drafting is you and a chat window with a good voice prompt.

Image generation. Still manual for me, using Nano Banana 2. I did write a helper for the thumbnails, but you could equally do them by hand or with a template in Canva.

If you strip out my two custom apps, what’s left is a database, a scheduler, a blogging platform, and an agent that talks to all three. Every one of those is an off-the-shelf product with a free or cheap tier.

How does it meet my goals?

At the start of this post I laid out my goals. How does this system meet them?

Goal: Controlled using AI agents

Storing the data in Airtable allows me to use Claude Cowork to control the social posting. I use Cowork’s big brother Claude Code to publish my blog posts but that’s only because that’s what I used to build the site in the first place.

Goal: Repeatable process

Most of the process is automated and the manual steps are the same every time. This means that when I find something that’s awkward or hit a new situation I hadn’t anticipated, I can change something in the process to make it better and that change will be in place for every subsequent post. Every fix I make sticks around for every post after it - small improvements add up fast

Goal: Uses AI to assist me, not take over

At every step of the process I can step in and make changes. Furthermore, I am the one who decides when to post - not an AI. The one exception to this is the post recycler. It runs automatically but I have control over what it posts and I have run it enough times that I trust it to do the right thing.

Trust is really important when it comes to getting the most out of AI. Building a repeatable, automated system like this means that I can progressively build trust in my agents each time I make an improvement. The more trust I have, the less I need to supervise the agent.

Goal: Makes managing the voice used in AI writing really easy

Having voices stored in Airtable instead of wrapping them up in skills as is commonly done, I can use anything to create a voice that is available everywhere in the system. My app that helps with writing definitely makes working with voices easier but there’s no reason not to do it using an agent or just writing voices by hand.

Goal: Schedules social posts in advance

Buffer is really simple and has good support for agents like Cowork.

Why it works

Four things, and none of them are about the technology.

I do the real thinking. AI helps with research and drafting. It doesn’t decide what I have an opinion about. That’s the difference between content worth reading and slop.

The voice definitions do a lot of heavy lifting. Writing down properly how I write - what I say, what I never say, how I open and close - took real effort once, and now every agent in the chain uses it. A small number of well-written artefacts like this will do more for your results than any amount of prompt fiddling.

I’m in the loop. Nothing publishes without me signing it off. Not because the agents are unreliable, but because it’s my name on it.

The shared database is the glue. Because the actual content - the articles, the voices, the images - lives in one place that everything can reach, I’m not locked into any of my choices. I can add a new agent when I think of a job for it. I can rip out a step and replace it when I find a better way. I can add new information to the pile, such as website traffic or how posts are performing, and point an agent at that too, without rebuilding anything I’ve already got.

Compare that to the alternative, where each tool holds its own copy of everything and nothing quite talks to anything else.

I used Cowork, but there’s nothing magic about that choice. The approach works with any platform that can connect to your everyday apps and run tasks on a schedule.

And it’ll get better. As I refine the voice and see more output, the results get more consistent. At some point I might let a few social posts go out without me reading them first, but only when I’ve watched it work often enough to trust it.

The point

Agents aren’t a technical speciality any more. Products like Cowork are built for people who have a business to run, not a codebase to maintain.

What I’ve described is a recipe. Central store of your content. Voice definitions so the output sounds like you. Agents for the mechanical steps. You reviewing before anything goes out. Swap Airtable for Google Sheets, Buffer for something else, LinkedIn for wherever your customers really are. The shape holds.

The reason I’d point you at agents rather than at another AI newsletter is simple: you learn what this technology can do by giving it a real job and watching what happens. Pick the task you’ve been avoiding. Mine was social media. Yours might be quoting, or chasing invoices, or writing up site visits.