Let an AI agent do the busywork

Chapters 1 to 5 describe a loop: write a batch, adapt it per network, schedule it, reply, read the log, adjust. About half of that loop is judgement — what to say, who to reply to, what the numbers mean. The other half is clerical: turning one idea into six versions, checking each against its network’s limits, scheduling, listing what went out, finding what failed. This chapter is about handing the clerical half to an AI agent you already use, and keeping the judgement.

It is optional. Nothing in the plan requires it. But if you already have Claude, ChatGPT or Cursor open all day, it removes most of the friction that makes people stop posting in week three.

How it works, in one paragraph

The Model Context Protocol (MCP) is an open standard that lets an AI client call tools exposed by a server. A scheduler that ships an MCP server is giving your agent a set of named actions — list my accounts, schedule this post, what failed — that the agent can call on your behalf. Claude supports adding a remote MCP server as a custom connector (on every plan; the Free plan is limited to one custom connector); ChatGPT supports connecting a remote server once developer mode is switched on in its settings; Cursor and Claude Code take a JSON or one-line config. The agent then reads the tool list and picks tools as you talk to it.

seenpaid’s MCP server exposes 50 tools at https://api.seenpaid.com/mcp. Every tool named below is a real one from that list; the MCP page has the full set and a config block per client, and the docs walk through the API key.

Connecting, once

1. In seenpaid, create an API key (Settings → API keys). It is scoped to one workspace.
2. Add the server to your client. Claude Code: claude mcp add --transport http seenpaid https://api.seenpaid.com/mcp --header "Authorization: Bearer sp_your_key". Claude Desktop, Cursor and ChatGPT: the config blocks on the MCP page.
3. Ask the agent: "Call list_accounts and tell me which social accounts are connected and whether any need reconnecting." If it answers with your handles, it is working.

Prompts for week 1: setup and the first posts

The agent cannot do OAuth for you — connecting a network is a human step, by design. It can tell you what is missing and hand you the link.

• "Use get_account_health. For each account that is not healthy, tell me why and give me the reconnect link."
• "Use get_platform_requirements for LinkedIn, X and Bluesky and summarise the caption limit, whether media is required, and whether links in the body are tracked, as a table."
• "Here are 20 ideas [paste the list from chapter 1]. For each, write a one-line hook in the mistake-list format. Do not schedule anything."
• "Take hook 3 and write a full plain-text post under 280 characters. Run validate_post against X and Bluesky and show me the result before doing anything else."

The last prompt is the habit to build: validate before schedule. The dry-run tool reports, per network, whether a post would publish and exactly why not, so a too-long caption is caught in the conversation rather than as a failed post.

Prompts for weeks 2–3: the batch, adapted and scheduled

This is where the agent earns its keep. The chapter 4 sitting — one idea, six versions, all scheduled — is one prompt.

• "Here is this week’s long-form post: [paste]. Write the X version (under 280), the Bluesky version (under 300) and the Threads version (under 500), keeping the first line identical. Validate each with validate_post. Then use schedule_post to schedule the LinkedIn original for Tuesday 09:00 Europe/Vilnius and the three short versions for the same day at 13:00. Show me the plan before you schedule."
• "Use get_next_slot to suggest the next open slot on LinkedIn, then schedule this post there."
• "Here are seven posts for next week as a numbered list, one per day, all for LinkedIn and X. Use bulk_schedule, starting Monday 09:00, one per day. Report the post ids."
• "Use list_posts with status scheduled and show me everything going out this week as a table: day, time, networks, first line."
• "Reschedule post [id] to Thursday at 10:00 using update_post. Do not change the caption."

Say the timezone every time. Leaving it out is an easy way for a scheduled post to land at the wrong hour, and a tool that takes an explicit timestamp is only as good as the timestamp you give it.

Prompts for every day: what went out, what failed

Twenty minutes a day was budgeted for replying. The two minutes before that are for a status check, and the agent does it well.

• "Use list_posts with status failed for the last 7 days. For each, use get_post and tell me the exact error per network and whether it is something I need to fix (like reconnecting) or something to retry."
• "Use list_notifications and summarise anything I should act on."
• "Use summarize_performance for the last 14 days and give me the short version: what went out, what got the most engagement, what failed, and one thing to do next."

The failed-post prompt is the one that matters most for the 30-day plan, because a silently failed post on day 12 is how a habit dies without anyone noticing. Networks expire tokens, change API versions and reject media formats; an agent that checks daily turns that into a reconnect link instead of a gap.

Prompts for week 4: reading results

If attribution is off (the default), the agent can read back engagement for the networks that report it and list your posts; the chapter 3 log stays yours. If you have turned attribution on and connected Stripe, the revenue tools become available:

• "Use get_top_posts for the last 30 days. For the top three, show me the caption’s first line, the network, clicks, sales and revenue."
• "Use get_dead_posts for the last 30 days — posts with clicks and no sales — and group them by network. Which landing page did they point at?"
• "Use get_channel_roi and tell me which network returns the most revenue per post, and which returns the least. Just the ranking and the numbers."
• "Use get_attribution_health and tell me what share of my Stripe revenue could be traced to a post, and what the product says is incomplete in the setup."
• "Take the top post from get_top_posts and use repost to republish it to the same network next Tuesday at 09:00, with the first line rewritten as a question. Show me the new caption first."

What to keep for yourself

A short list of what is worth keeping, and why:

• The ideas. An agent can shape twenty ideas into a hundred posts; it cannot know the thing you learned last Tuesday that your buyer needs to hear. The list in chapter 1 stays hand-written.
• The replies. Automated replies are recognisable and they cost the one thing that makes a small account grow, which is that the person answering is the person who built the thing.
• The publish decision. Every prompt above asks the agent to show the plan before scheduling. Keep that. It costs one message and it catches the wrong-timezone, wrong-account and wrong-link mistakes that no validator can.
• The judgement about results. The agent can rank posts; deciding whether one sale is a pattern is chapter 5’s job, and it is yours.

A full week, as one conversation

Putting it together, a Monday sitting with the agent looks like this: check health and failed posts; paste this week’s ideas and get hooks back; pick three; get the long versions drafted, edit them yourself; get the per-network versions, validated; review the plan; schedule with bulk_schedule; ask for the scheduled-this-week table as a sanity check. Roughly forty minutes, most of it your edits. Then the rest of the week is replying, and a two-minute status check each morning. That is the 30-day plan with the clerical half removed, which is the version most likely to still be running in month six.