How to Build a Social Media Agent with Claude
Facts checked on October 17, 2026
A social media agent built with Claude is three things: a written brief that tells Claude how your brand posts, a connector to a scheduling tool that can draft, schedule and read your calendar, and a person who approves before anything goes live. Claude’s own features cover all three, and none of them needs code. The five steps below set it up and run it once a week.
What a social media agent is#
Here “agent” is not a separate product. It is Claude doing a repeated job with three pieces in place.
| Piece | What it does | Where it lives in Claude |
|---|---|---|
| A brief | Your voice, your networks, your rules | A project or a skill |
| A connector | Sees your accounts, creates and schedules posts, reads the calendar | Customize, then Connectors |
| An approval step | A person reads each post before it goes out | Claude’s prompts, and the scheduling tool’s own |
Claude writes the posts. The scheduling tool holds your permission to post on each network and sends each post at its time. The agents page shows how the two fit together.
Step 1: Write the brief#
Put the brief where Claude reads it every time. Anthropic offers two places.
A project. Anthropic’s article on projects calls them “self-contained workspaces with their own chat histories and knowledge bases”, each with its own instructions. Projects are on every plan, up to five on Free. Paste the brief into the project instructions.
A skill. A folder of instructions that Claude opens when a request matches it. Skills are on every plan once the code execution setting is on, and you upload one as a ZIP file under Customize, then Skills. See what goes in an agent skill for social media posting.
Start with a project. The same text can become a skill later.
A brief to copy and adapt:
You are the social media assistant for [Brand], a [what you sell] for [who buys it].
Voice: friendly and plain. Short sentences. No hype. At most one emoji per post.
Networks: LinkedIn (company page), Instagram, Bluesky. One version per network.
When I ask you to plan the week:
1. Read the calendar for the next 7 days and list the empty days.
2. Draft one post for each empty day from my notes.
3. Show me every draft with its network, date and time.
4. Schedule only the drafts I say yes to.
Never mention prices, discounts or customers without asking me.
If the scheduling tool shows an error, tell me and stop.Step 2: Connect a scheduling tool#
MCP is the standard way Claude and ChatGPT plug into other apps. A scheduling tool that offers it gives you an address, and in Claude that address becomes a connector.
Anthropic’s connector guide says custom connectors are on every plan, and “Free users are limited to one custom connector.” On Pro or Max:
- Go to Customize, then Connectors.
- Click “+ Add”, then “Add custom connector”, and name it.
- Paste the address the scheduling tool gives you and click “Continue”.
- Sign in to the scheduling tool. Claude acts for you “without Claude ever seeing your actual password”.
- In a chat, click “+” at the lower left, open Connectors and switch it on.
On Team and Enterprise, an Owner adds the connector first, then each member connects their own account.
The connector should let Claude see your accounts, create and schedule posts, and read the calendar. Without the calendar, Claude cannot see what is already planned.
If Claude ignores the connector, set it to “Always available” in its tool access setting. The full setup, Claude Code included, is in how MCP connectors let Claude use a scheduler.
Postnavo is built for this step. At launch, its connector will let Claude list your channels, create and schedule posts, and read your calendar and queue across ten networks. The setup is on the Postnavo page for Claude.
Step 3: Give it your calendar and your voice#
A brief describes the voice. Material shows it.
- Past posts. Add five to ten of your best posts as files in the project. Claude copies an example more closely than it follows an adjective.
- A content plan. Themes, launches and dates to avoid for the month.
- The live calendar. Ask Claude to read it through the connector before it drafts, so it fills gaps instead of stacking posts.
Step 4: Keep a person approving#
Claude’s own prompts. Claude asks before it uses a connector’s tool. Anthropic’s advice is to click “Allow always” only “when using a server and tool that you trust to run unsupervised.” Keep the prompt for anything that schedules. On Team and Enterprise, an owner can set each action to Always allow, Needs approval or Blocked.
One exception: during Research, “Claude can invoke tools from your connectors automatically without further approval.” Switch the scheduling connector off before you run Research.
The scheduling tool’s own approval. “Allow always” removes Claude’s prompt for that tool, so a second hold is safer. Postnavo is launching with an approval setting: with it on, a post Claude creates will wait in Postnavo for your OK. See the human-in-the-loop checklist for AI social posts.
Step 5: Run it every week#
By hand. Open the project every Monday and write “Plan next week”. This works on every plan.
As a scheduled task. Anthropic’s page on scheduled tasks says they “are available on all paid plans (Pro, Max, Team, Enterprise)”. You describe the job once and pick how often it runs: hourly, daily, weekly, on weekdays, or manually. Tasks run “even when your computer is asleep”, and can use your connectors and skills.
Whether a task reads a project’s instructions is not documented, so for an unattended run put the brief in a skill. Anthropic also lists an approval mode among a task’s settings, without saying on that page what each mode allows. So word the task to draft, not publish: “Every Monday at 9 AM, read my calendar for next week, draft a post for each empty day following my social media skill, and leave them for my approval.”
Once scheduled, each post waits in the scheduling tool’s calendar until its time. Claude does not need to be open.
What does not work#
- Claude with no connector. It writes the posts, and you copy each one in yourself.
- “Allow always” with no approval in the scheduler. Nothing then stands between a draft and your followers.
- A brief made of adjectives. “Be engaging” gives Claude nothing to copy. Real posts do.
- A connector from an address you found somewhere. Anthropic warns that a malicious one “may include hidden instructions that try to make Claude perform unintended actions.” Add only the address the tool’s maker publishes.
FAQ#
Do I need to code to build a social media agent with Claude?#
No. The brief is plain text, and a connector is added in Claude’s settings with an address and a sign-in.
Which Claude plan do I need?#
Projects, skills and one custom connector are on the Free plan. Scheduled tasks need Pro, Max, Team or Enterprise.
Can the agent post without asking me?#
Only if you let it. Claude asks before it uses a connector’s tool unless you clicked “Allow always”, and Research skips that prompt. A scheduling tool with its own approval setting holds the post either way.