Most of the work in running a social calendar is the same few jobs, week after week. Fill the gaps, look for what is about to go wrong, find an old post worth another go, work out why something failed, see what needs you this morning.
Orbitwrite has an MCP server and a Claude connector, and five of its built-in prompts cover those jobs. Here are the words to type for each, what Claude reads or changes in your workspace, and where it stops and waits for you.
Connect Claude first
Setup is short and the ChatGPT post already covers it in detail. In Claude, add a custom connector, paste the Orbitwrite URL, sign in, pick the workspace and press Allow.
https://mcp.orbitwrite.com/mcp
Or let Claude walk you through it. Open this prompt in Claude, or paste it into Claude Code, and it fetches Orbitwrite's setup instructions and follows the steps for wherever it is running:
Fetch and execute the appropriate instructions to set me up for Orbitwrite from https://orbitwrite.com/agent-setup/prompt.md
Settings, MCP in Orbitwrite has a snippet for each client, Claude Code included. Start with a question that checks the connection:
Which Orbitwrite workspace am I in, and what is the time there?
Claude calls whoami, which returns the workspace's timezone and the current time in it. Every time in the prompts below is read and written in that zone.
Prompts, or just words
Orbitwrite registers these workflows as MCP prompts. In a client that lists MCP prompts, they show up by name, and the host can turn them into slash commands. Which names appear depends on what you consented to: a prompt is only offered when every tool it uses is available to you, so a read-only connection will not list the ones that change things.
You do not need the prompt picker. Each section gives a plain sentence that asks for the same job, and the picker version adds a fixed set of steps, in order, with the points where it must ask you first.
1. Plan a week
Plan 5 posts for the coming week about our autumn launch.
What it reads: the week overview for the seven days from today (open posting slots, your channels and the workspace timezone), your brand voice, and your tags.
What you see back: a table of slot, channels, angle and tags, and a question. Nothing is saved until you confirm or edit the table.
What it changes: once you agree, each post is saved as a draft with each channel's own copy and placed on its slot as a planned post. Planned posts hold the slot on the calendar without being scheduled, so the worker never publishes one. Claude then asks who should review. Name people and it asks for an approval request on each draft, blocking unless you say it is advisory.
It ends with each post's id, slot and reviewers. The prompt's own last line is that nothing has been scheduled, and that is the behaviour you should expect.
2. Check the calendar
Check what goes out in the next 3 days.
What it reads: the week overview for failed publishes, posts waiting for approval and changes requested, then the scheduled posts in the window. For any scheduled post with an issue listed, or a post score grade of fair or needs work, it scores the post again.
What you see back: problems by day, soonest first, with the channels and the time, then one suggested fix for each and a question about which to apply.
What it changes: nothing until you answer. It is a read-only check.
The point is to find the problem post while there is still time to fix it.
3. Repost a winner
Find a post from the last 90 days worth reposting and prepare a fresh version.
What it reads: your top posts by performance for that window. It picks three that are evergreen, meaning not tied to a date or an event, and says why each worked. A metric a network did not report is treated as missing, not as zero.
What you see back: the three candidates, and a question about which to repost.
What it changes: after you choose, it duplicates the post, rewrites the opening line while keeping what made it work, and saves the result as a draft. It scores the draft and fixes anything that fails, then offers the first free slot for those channels. Only if you agree does it schedule the copy.
4. Recover failed posts
Find posts that failed to publish in the last two weeks and get them out.
What it reads: the failed publishes from the week overview, then each channel's error, and the post's activity when the error does not explain itself.
What you see back: each failed channel sorted into one of three groups, with the reason in plain words:
- Retryable, such as a timeout, a rate limit or an outage at the network.
- Needs a reconnect, with the name of the channel to reconnect in Orbitwrite.
- Needs an edit, with what the network refused.
What it changes: after you agree, it retries the retryable channels one at a time and reads the publish status again to confirm. Reconnecting a channel is a browser step, so the prompt tells you which one and leaves it to you. Orbitwrite does not offer a connect or disconnect tool over MCP.
A retry sends the stored content again, so a channel that already published is refused rather than posted twice.
5. Morning check
Give me the morning check.
What it reads: your workspace's current time, the activity feed since the same time yesterday, the unread conversation count, and today's schedule.
What you see back: five lines or fewer. What published or failed since yesterday, the approvals waiting on you, the unread inbox count, and what goes out today at which times.
What it changes: nothing. It is read-only.
Where you stay in charge
A prompt is only instructions to the model. The rules sit behind the tools, and they are the same ones that apply to everything else in your workspace.
- You choose what Claude can reach. The consent screen offers two bundles, one for reading and one for writing. Claude works under your role in the workspace, limited to what you approved, so a connection that can only read has no tools to plan, schedule or retry.
- Some things are out of reach whatever you approve. API keys, billing, your team, and the logins behind your channels are not available to a connector.
- Prompts that publish or schedule ask you first. The wording in each one tells Claude to stop and wait for your answer before it schedules, retries or replies.
- Approvals still apply where they are on. On a plan with approvals, a member who cannot publish gets their scheduled posts held behind an approval request, and that holds for a post Claude schedules for them as well. Approvals are part of the Advanced plan, so on other plans there is nothing to hold.
- You can cut it off. Open Settings, then MCP, and remove the app under Connected apps. Access ends straight away.
The prompt wording asking for your go-ahead is a request to the model. The permission checks and the approval gate are what enforce it.
From ChatGPT, Grok, OpenClaw and Muse
ChatGPT connects in the same way, with the same URL and a sign-in, and the ChatGPT post walks through it. Grok connects with an API key: open grok.com/connectors, choose New Connector, then Custom, paste the same URL and give the key when it asks. OpenClaw and Muse connect with the same URL and an API key sent as a header. Settings, MCP creates the key and shows the steps and connection details for each. Those connections work with the sentence versions of the prompts above. Whether a client lists MCP prompts is up to the client.
There are five more prompts that I have not covered here: a weekly review, filling the week with ready drafts, triaging the inbox, drafting from a link, and planning a campaign. They work the same way.
Everything above is on the MCP page. To try the first one, connect Claude and ask it to plan the coming week.