Advanced: Run a Content Loop With an Agent, From Ideas to Weekly Review

What you get

If you run several platforms on your own, most of the time goes to chores outside of writing: finding topics every day, reformatting posts for each platform, tracking post times, and then digging through dashboards for numbers. This guide shows you how to borrow the structure of a commercial tool and use Claude Code or Codex, agents that can read and write files, to set up a content loop on your own computer. It collects material, picks topics, rewrites each draft for each platform, schedules posts, and feeds the review back into topic picking. No coding needed. The article includes 5 prompts you can copy and paste directly.

Be clear about the split: the agent handles collecting, scoring, drafting, and reports. The publish button and the final look stay with you. Automation saves repetitive work. If something goes wrong with your account, the responsibility is still yours.

Why split it into five modules

This split comes from a social media automation tool called Posteady. It breaks the process into five stages: Collector, Picker, Creator, Scheduler, and Reviewer. At the end, it sends the content that performed well back into the topic pool. Splitting it up means each stage can be checked and replaced on its own:

• Each stage hands off one thing: a source library, a topic list, drafts, a posting calendar, or a weekly report. When something goes wrong, look at what that stage produced.
• You guard only two checkpoints: approving topics and a look before publishing. The agent handles everything else, so you can review it in ten minutes a day.
• The loop gets sharper over time: the review writes down which content people actually read and feeds that into your topic criteria. The next round of topic picking then rests on evidence, not gut feel.

First, decide: buy off the shelf or build your own

	Buy a ready-made tool (Posteady as the example)	Build your own (the method in this article)
Cost	List price is $29 to $149 a month, with a 7-day trial	The subscription or usage of the agent itself
Setup	Connect your accounts and start. The interface is friendly	Takes an afternoon or two to get the files and prompts in order
Publishing	Connects to your platforms and posts on a schedule	Use each platform's built-in scheduled posting, or connect a separate publishing tool
How far you can change it	Limited to the features the product offers	You decide the topic criteria, brand voice, and report format
Data	On the vendor's servers	Sources and drafts stay in your own folder

If you want to start today without fuss, try a ready-made tool. If you want to keep your topic criteria and writing style in your own hands, read on.

Step 1: Set up a content workspace

Create a folder on your computer, for example content-loop, and send this to your agent:

In the current directory, set up a content workspace for me with the structure below. Each file should start as a template with clear comments:
- sources.md: the sources I follow (RSS, YouTube channels, GitHub projects, newsletters), one per line, each with a sentence on why I follow it
- voice.md: my brand voice (who I am, who I write for, phrases I often use, things I never say, and 3 older pieces I'm happy with as samples)
- inbox/: material collected each day, one file per day
- picks.md: the topic list
- drafts/: drafts sorted by platform
- calendar.md: the posting calendar (date, platform, draft path, status)
- reports/: weekly reviews
After setting it up, list the directory structure. Then ask me what to fill in sources.md and voice.md, one question at a time.

voice.md deserves the most time. How much the agent sounds like you depends on the samples and the "never say" list in this file.

Step 2: Collect sources (Collector)

For sources, prefer public RSS feeds, official APIs, or newsletters you subscribe to yourself. Don't have the agent scrape pages that require login, and don't bulk-scrape social platforms. That usually breaks platform terms and puts your account at risk.

Run it once a day:

Read each source in sources.md and collect new content from the past 24 hours. Write it to inbox/<today's date>.md.
For each item, write: the title, the link, the source, and one sentence on what it covers. If a source can't be fetched, mark it "failed" and write the reason. Don't make up content.

Claude Code and Codex can both set this as a task that runs on a daily schedule. If you don't want to set a schedule, you can also run it by hand each morning.

Step 3: Pick topics (Picker)

With a lot of material, the hard part is choosing. The prompt below turns picking into scoring, so you only need to look at the top few:

Item 2 does the most to prevent sameness. When everyone is sharing the same news and you have no view of your own, what you write is just the hundredth version of it.

Step 4: Write one draft for many platforms (Creator)

Pick one topic from picks.md, then send:

Make topic number [number] from picks.md. First read voice.md, then write one version per platform under drafts/:
- X: one post of up to 280 characters, plus an optional 3-post thread version
- LinkedIn: about 150 words, with the conclusion in the first sentence
- Instagram carousel: 5 to 7 slides, each with a one-line headline and two lines of explanation
At the end of each version, list the facts and source links you used. Mark any fact you're unsure of as "to verify". Don't make up numbers. Stop after writing and wait for my edits.

The platforms differ in more than word count. X needs a hook in one sentence, LinkedIn needs the conclusion first, and a carousel covers one point per slide. Having the agent write separately for each platform works much better than pasting one draft everywhere.

Step 5: Schedule and publish (Scheduler)

After you edit a draft and confirm it, have the agent update calendar.md with the date, platform, draft path, and status for each item. Keep the act of clicking publish for yourself. The scheduled-posting feature built into each platform is enough.

Once this workflow has run steadily for a few weeks, you can consider connecting a publishing tool. At that point, give the agent a dedicated account with minimum permissions, and keep a manual check before anything goes out.

Step 6: Weekly review (Reviewer)

Each week, export data from each platform's dashboard (most platforms can export CSV), put it in reports/, then send:

Read this week's exported data in reports/ and calendar.md, then write a weekly report to reports/<this week's date>.md:
1. The top 3 and bottom 3 posts, each with its data and the reason you think it happened (state that this is a guess)
2. Which platforms, formats, and time slots performed better
3. 3 items for "do more next time" and 3 for "stop doing". These will be read the next time you pick topics.
4. For the best-performing post, give 2 topics that could be done again from a different angle, and add them to picks.md
Each platform measures data differently, so don't add the numbers together and compare them directly.

Item 3 is the thread that closes the loop. The conclusions the review writes down get read the next time you pick topics.

Notes

• Don't scrape what you shouldn't: use only public RSS, official APIs, and content you subscribe to yourself. Bulk-scraping social platforms or logging into someone else's account may break platform terms and risks getting your account banned.
• Don't let it publish on its own: auto-publishing is the easiest way to go wrong, so keep a human look between drafting and publishing.
• Credit sources when you paraphrase: if you use someone else's views or data, name the source in the draft. Don't rewrite someone else's content and call it original.
• Run it for two weeks before adding features: first get collecting, picking, writing, and reviewing running steadily every day. Then consider competitor monitoring and auto-publishing.

References

• Posteady website
• Claude Code docs
• OpenAI Codex docs