How-To Guide
How do you build a safe Claude workflow for social media?
A five-stage Claude workflow for social media, from research to report, with the review and approval gate between each stage, built from vendor rules.
Build it as a workflow, not an autonomous agent: a fixed five-stage path that researches, drafts, prepares, publishes and reports, with a check at every stage, and a person or an account role holding the release decision. Anthropic defines a workflow as LLMs and tools orchestrated through predefined code paths, which is what makes the path reviewable. Claude does not publish by itself, so a connected tool performs the action, and the strongest place for the final yes is the publishing account, not the prompt. Superdirector draws the same line in its own product: its MCP connection reads the Canvas and saves drafts, and the connected AI cannot post. Posting is turned on for each account on the Canvas.
- Use a workflow when the job repeats and the steps are known; keep an agent for open-ended work you cannot pre-plan.
- Put a review at the end of research, drafting and preparation, and put the release decision in the account that publishes.
- Scope the connector to the tools the job needs, because a connector can change data as well as read it.
- Require a report from every run; a quiet scheduled task is not proof that anything published.
The short answer: a shorter path, not a smarter model
If you search for Claude workflows for social media, most results sell the model's reach: a list of jobs, crowds of agents, a month of posts while you sleep. The safer question is narrower. What is the shortest path from an idea to a published post in which every step can be inspected, and where does a person actually stop it?
Anthropic's own guide to building effective agents gives the frame. A workflow is a system where models and tools are orchestrated through predefined code paths. An agent is a system where the model directs its own processes and tool use. A repeating social job has known steps, so a workflow buys you predictability and a place to put a check. That is the trade this page makes.
The second half of the answer is about publishing. Claude drafts but does not publish on its own; a connected tool does. A competitor page states it in one line, that Claude on its own drafts but cannot publish, and it is right about that even where its numbers are its own. The useful design work is deciding which connected tool may act, what it may touch, and who releases the final post.
Superdirector's MCP overview shows one deliberate version of that boundary. The connected AI can read the Canvas, save drafts and start work after showing a credit estimate, but your AI cannot post. Posting stays on the Canvas, where you turn it on for each account. Everything below can be used with any connector, and this product choice is one worked example of the boundary, and other connectors draw it differently.
Original diagram from Anthropic's workflow guidance, Buffer's approval documentation and Superdirector's MCP overview, all stored November 18, 2026. The gates are advice about where a person or an account can stop the path, not a product default.
Make it work for your brand
Find it. Remake it. Post it. Superdirector turns the viral formats in your niche into your brand's carousels, X posts and video scripts, and keeps the calendar full.
Workflow or agent: the choice that sets the risk
The words are used loosely online, so it helps to keep Anthropic's distinction in view. A workflow runs a sequence you can read before it runs. An agent decides its own next step while it runs. Neither is better in general. For a weekly posting job, the sequence is knowable, and a knowable sequence is the thing you can approve.
Anthropic also advises finding the simplest solution possible and only increasing complexity when needed. That sentence is the practical counterweight to the dynamic, parallel, many-subagent workflows that dominate this keyword. A brand that has not yet run one reliable post through one connector will not be saved by a fan-out of parallel workers. Start with the simple chain, and add branches only when a real failure shows you need them.
| Decision | Workflow | Agent |
|---|---|---|
| Who chooses the next step | A path written in advance | The model, while it runs |
| Best for | A repeated job with known steps | Open-ended work with unknown steps |
| What a person can review | The whole path, before it runs | Outputs and checkpoints as they arrive |
| Failure to watch | A stale or wrong step in the path | Compounding errors across many steps |
The definition on the competitor page is compatible with this: a Claude workflow is a repeatable job you hand to Claude, and Claude plans each run while the job stays repeatable. What separates the two is what you can inspect before anything is published.
The five stages, and the gate between each
The workflow below has five stages. Three of them end with a person's decision, one ends in an account-level approval, and the last one produces evidence. The stages are deliberately boring. A stage that needs its own explanation is usually two jobs wearing one name.
- Define the job and the done test. Output: one paragraph a new teammate could check, naming the account, format, sources, link and time.
- Research with the sources attached. Gate: a person confirms every source is dated and reachable before a claim moves on.
- Draft in the brand voice. Gate: a person reads the draft against the done test, with the sources still in view.
- Prepare and publish through a connector. Gate: the post is created on hold, previewed, and released by the account's approval role or by the exact preview.
- Report and audit the run. Output: a success or failure result, and the destination checked for the post.
Two stages carry most of the risk. Research can import a wrong fact that the rest of the chain faithfully repeats. Publishing can act before anyone has read the result. That is why the checks at those two points are worth more than a dozen extra prompts elsewhere. A checker agent can assist, but it is still a model. A human yes, or an account role that withholds publication, is a different kind of control.
Stage 1: Define the job and the done test
Write the job as one sentence and the done test as one paragraph. The job names the audience and the outcome. The done test names the account, the format, the source rule, the link rule and the time window. If those five things are not written down, the rest of the workflow has nothing to check against, and every later review becomes a taste argument.
The competitor page has a sharp line for this, that if you cannot say what done looks like, you do not have a loop, you have a wish. Anthropic makes the same point in its architecture language: agents can pause for human feedback at checkpoints or when encountering blockers, but a checkpoint only works if the person knows what passing looks like.
Keep the first job small enough to fail cheaply. One account, one format, one week. A workflow that also answers comments, reads direct messages and reports on three networks is three workflows and a reporting job, and bundling them hides which part went wrong.
Stage 2: Research with the sources attached
The research stage reads pages, posts, analytics or a media library, depending on the tools you grant. That is a permissions decision before it is a prompt decision. Anthropic's connector documentation is blunt about the stakes: connecting Claude to an external service grants it the ability to access and potentially modify data within that service, based on your permissions. It tells users to connect only to trusted servers and to review requested permissions carefully.
Two concrete limits shape the setup. On Claude, free users can add one custom connector, so a multi-tool research stage is a paid-plan decision from the start. And a connector exposes the tools its vendor chose, so review the list and disable what the job does not need. If the research stage only reads, it should not hold an upload or a schedule tool.
The gate is a person checking provenance, not vibes. For each claim that will appear in the post, the research output should name the page, the date and the sentence that supports it. A number with no stored page behind it is not evidence, and a page that cannot be reached again is not much better. This is the cheapest place to stop a wrong fact: before it has been rewritten into your brand's voice.
Our own connector is deliberately narrow here. Its documentation says the AI sees only what the Canvas has saved and does not see your computer files or private accounts. Scope is a product feature you can inspect, and it is worth asking any connector vendor to state their scope this plainly.
Stage 3: Draft in the brand voice
Put the voice, the network rules and the done test in a skill or a standing instruction, so they do not have to be repeated in every prompt. PostFast's guide makes the boundary clear: a skill on its own cannot post anything. It shapes the output and makes the review shorter, and it cannot grant access.
Draft one version per network. Do not paste one text everywhere. The tool usually knows the local limits, and the draft stage is where those limits should surface. If a connector rejects a post for length or a missing image, fix the post. Do not hunt for a way around the check.
The gate is a read against the done test. Does the draft name the right account, use the agreed format, carry the approved sources, place the link where the rule says, and fit the time window? A person who can answer those five questions in a minute is doing a real review. A person scrolling past a preview is not.
Keep the sources beside the draft until the post is published. Once the copy is polished it becomes easy to forget which sentence came from which page, and that is exactly when the review stops being a check and becomes a proofread.
Stage 4: Prepare and publish through a connector
Separate two questions that are often confused. May this tool act at all? And has a person approved this exact post? The first is a permission question, answered by the connector's scopes and the account you signed in with. The second is an approval question, and it is strongest when it lives in the publishing account.
Buffer's approval documentation is a clean example of the second kind. A team member with the Needs Approval role always leaves a draft awaiting approval, whether the post came from the dashboard, the mobile apps, the Buffer API or the MCP server, and an approver is notified by email. Approval on Buffer's Team plan is a workspace permission, so it survives a prompt rewrite, a model change or a new operator. Our MCP overview takes a related route: the connected AI has no posting tool, and posting is switched on per account on the Canvas.
Where the account has no approval role, the hold-and-preview pattern is the fallback. A connector vendor describes it directly: the post is held until you approve it, and Claude creates posts on hold with a date and time, up to a year ahead, and schedules them once you approve the preview. That is a person saying yes to exact bytes, which is stronger than a model saying the work is done.
Then check the Claude side of the connection. Claude Code's MCP documentation shows that a project-scoped server sits as pending approval until a person runs the tool interactively to approve it, and that a server can be blocked with a settings key. Those are connection-level controls. They decide whether the tool exists in the session, not whether a particular post is good.
Original diagram built from Anthropic's custom connector page, Buffer's approving-draft-posts article and Superdirector's MCP overview, stored November 18, 2026. The ordering is a reading of those sources, not a claim that every connector implements all three layers.
Stage 5: Report and audit the run
The last stage has two jobs. Prove that the run happened, and judge whether it worked. Keep them apart. A scheduled task can finish cleanly and still have published nothing, and a post can go live and still be the wrong post.
Require a result in both directions. A competitor page puts it well: silence is not proof that the automation worked. The run should return success or failure, the account, the destination and the time, and a failed run should say why. When a scheduled workflow publishes nothing for several days, the operator needs an alert, not an inference from an empty calendar.
Read the destination yourself. The connector's success message is a claim about an API call, and the page on the network is the evidence. This is the same discipline Anthropic describes when it says agents can pause for human feedback at checkpoints: the pause matters because someone looks at ground truth, not because a status turned green.
Keep the evidence with the post. A saved result that names the account, the exact text, the time and the destination page is enough to answer the question a client asks later. The workflow is only auditable if that record survives the chat window.
Original diagram from Anthropic's guidance on ground truth and human checkpoints, a vendor page stating that silence is not proof the automation worked, and Superdirector's MCP overview, stored November 18, 2026.
The permissions checklist before you switch it on
Run this once for the account, before the first unattended run. It takes less time than a single mistaken post.
- Name every connector in the chain and the company that runs it. Remove any you cannot identify.
- Review the requested scopes at sign-in and after every update. Anthropic's guidance is to review requested permissions carefully.
- Disable tools the job does not use, especially upload, delete and schedule actions in a read-only stage.
- Set the publishing account's approval role or hold setting, and test it with one draft created through the connector.
- Write the done test where the operator can see it, and require a run report that names success and failure.
If the chain includes Claude Code, two more settings belong on the list. A project-scoped MCP server stays pending approval until a person approves it, and a server can be blocked for every session through the disabled-server setting. Check both before a headless run, because a server that is present and approved will act without anyone watching.
Treat anything in research preview as moving. Anthropic currently marks cloud scheduled tasks, formerly routines, as a research preview and says behaviour, limits and the API surface may change. A workflow that depends on a preview feature should have a manual fallback, so a quiet change does not stop the account.
Start with one stage and one account
The tempting first project is the whole chain. The useful first project is one stage that already works, repeated with a check. If drafting is your bottleneck, automate research and drafting, and keep publishing manual until the drafts pass your done test several times. If publishing is the bottleneck, keep drafting manual and put the connector and the approval role in place first.
Budget for three kinds of cost before you promise anything. Your Claude plan pays for the model work. A connector vendor bills on its own plan. A product such as Superdirector bills on its own terms: Growth is $199 a month for one AI-run operated account, and the pricing page's rows include two accounts at that price, with each extra account $99 a month. Starter is $29 a month for work inside Claude or ChatGPT. None of those numbers includes the time a person spends at the review gates, which is the cost that decides whether the workflow is worth running.
Then let it earn more rope. Add a network, add a format, add an unattended run, one change at a time, and keep the report after each change. A workflow you can audit is worth more than a workflow that can do more.