LinkedIn Research Post

You help the user research a topic and publish a LinkedIn post backed by multi-source research. The full pipeline: Research -> Analyze -> Draft -> Review -> Publish.

Prerequisites - VERIFY BEFORE STARTING

Before launching any agents, run these checks:

1. AnySite MCP canary check: Call mcp__anysite-mcp__duckduckgo_search with query "test" and count 1. If it returns results, AnySite is connected. If it errors, tell the user:

• Claude Desktop: Customize > Connectors > AnySite > Connect
• CLI: claude mcp add anysite -s user -- npx -y @anthropic/anysite-mcp
• Signup: anysite.io - promo code BAYRAMMCP for 30 days free

DO NOT launch research agents until this check passes.

2. Voice profile: Check for voice/style in this order:

1. SOUL.md in project root or home directory
2. user-profile.md in project root
3. Memory files via recall
4. If nothing found, ask the user: "How would you describe your writing style in 1 sentence?"

3. Unipile MCP (optional - check only at publish time):

• Signup: unipile.com - 7-day free trial, no credit card required
• Config: claude mcp add-json unipile '{"type":"http","url":"https://developer.unipile.com/mcp","headers":{"X-API-KEY":"YOUR_API_KEY"}}'

Workflow

Step 1: Get the Topic

If the topic was provided via $ARGUMENTS, use it directly.

Otherwise ask: "What topic should your LinkedIn post be about?"

If too broad, narrow it:

• "What specific angle or insight do you want to share?"
• "Who is your audience on LinkedIn?"

Step 2: Research (2-3 min)

Launch 3 parallel sub-agents. Each agent prompt MUST include the exact MCP tool names below. Do NOT say "use AnySite MCP" - agents don't know what that means.

CRITICAL: Include these exact instructions in each agent prompt:

Sub-agent 1 - LinkedIn perspective:

Research "[TOPIC]" on LinkedIn.

You MUST use these exact MCP tools (not WebSearch, not WebFetch):
- mcp__anysite-mcp__search_linkedin_posts(keywords="[TOPIC]", count=10, date_posted="past-week")
- mcp__anysite-mcp__get_linkedin_post_comments(urn="[post_urn]", count=5) for top posts

Find the 5 most engaged posts. For each note:
- Author name and headline
- Post text (first 200 chars)
- Reaction count and comment count
- The angle/take they used

Summarize: what angles get traction, what's overdone, what's missing.

PRIORITY: Try mcp__anysite-mcp tools first. If they fail, fall back to WebSearch/WebFetch.
At the end of your report, note which tools you actually used (MCP vs fallback).

Sub-agent 2 - Community perspective:

Research "[TOPIC]" on Reddit and Twitter.

You MUST use these exact MCP tools (not WebSearch, not WebFetch):
- mcp__anysite-mcp__search_reddit_posts(query="[TOPIC]", count=10, sort="top", time_filter="month")
- mcp__anysite-mcp__search_twitter_posts(query="[TOPIC]", count=10)
- mcp__anysite-mcp__get_reddit_post_comments(post_url="[url]") for top 2-3 posts

Find honest opinions, complaints, praised solutions.
Note where community sentiment DIFFERS from LinkedIn's polished narrative.

PRIORITY: Try mcp__anysite-mcp tools first. If they fail, fall back to WebSearch/WebFetch.
At the end of your report, note which tools you actually used (MCP vs fallback).

Sub-agent 3 - Deep content:

Research "[TOPIC]" on YouTube and the web.

You MUST use these exact MCP tools (not WebSearch, not WebFetch):
- mcp__anysite-mcp__search_youtube_videos(query="[TOPIC]", count=5)
- mcp__anysite-mcp__get_youtube_video_subtitles(video="[video_id]") for the top result
- mcp__anysite-mcp__duckduckgo_search(query="[TOPIC]", count=5) for web articles
- mcp__anysite-mcp__parse_webpage(url="[article_url]") to read top articles

Extract key frameworks, data points, expert claims.

PRIORITY: Try mcp__anysite-mcp tools first. If they fail, fall back to WebSearch/WebFetch.
At the end of your report, note which tools you actually used (MCP vs fallback).

After launching agents: Check their output within 15-20 seconds to verify they are using the MCP tools, not WebSearch/WebFetch. If any agent is using the wrong tools, send a corrective message immediately via SendMessage with the exact tool names.

Synthesis: Present to the user:

• Key themes across platforms (note which platform said what)
• "Different Truths": LinkedIn=aspirational, Reddit=honest, Twitter=reactive, YouTube=deep
• The content gap: what's trending vs what's missing
• 2-3 potential angles for their post, each mapped to a content template

Step 3: Draft (1-2 min)

Ask the user which angle and template they prefer:

A: "Lesson Learned" - Hook -> Mistake -> Learning -> 3-5 Takeaways -> Question B: "Curated Insight" - I found X -> 3 insights with context -> My take -> What are you seeing? C: "How I Do X" - Problem everyone faces -> My setup (3 steps) -> Result with numbers

Read the user's voice profile (SOUL.md, user-profile.md, or memory) to match their tone.

Draft the post following these LinkedIn best practices:

• Under 1300 characters (sweet spot for engagement)
• Opening hook in the first line (before the "see more" fold)
• No hashtags in the body text
• 3 hashtags maximum at the very end
• Line breaks between paragraphs for readability
• End with a question that invites discussion
• Use specific data points from the research (not vague claims)

Step 4: Review

Present the draft to the user. Ask:

• "Does this sound like YOU? What would you change?"
• "Any facts to verify or personal details to add?"

If the user wants a second opinion, offer to run /council:

/council "Should I publish this LinkedIn post? Is the angle compelling? What would make it stronger?"

Incorporate feedback. Iterate until the user is satisfied.

Step 5: Publish

When the user approves, publish via Unipile MCP.

Unipile MCP is an OpenAPI proxy with these tools:

• mcp__unipile__search-endpoints - find the right API endpoint
• mcp__unipile__get-endpoint - get endpoint details (params, auth)
• mcp__unipile__execute-request - execute the API call via HAR format

To publish a LinkedIn post:

1. First, get the user's Unipile base URL and API key. Ask them for:
   • Base URL (e.g., https://api1.unipile.com:13118) - subdomain and port are user-specific
   • API Key - from their Unipile dashboard
2. List their LinkedIn accounts to get the account_id:

   mcp__unipile__execute-request({
     "harRequest": {
       "method": "GET",
       "url": "https://{subdomain}.unipile.com:{port}/api/v1/accounts",
       "headers": [{"name": "X-API-KEY", "value": "{their_key}"}]
     }
   })

3. Create the post using multipart/form-data (CRITICAL - JSON body does NOT work):

   mcp__unipile__execute-request({
     "harRequest": {
       "method": "POST",
       "url": "https://{subdomain}.unipile.com:{port}/api/v1/posts",
       "headers": [
         {"name": "X-API-KEY", "value": "{their_key}"},
         {"name": "Content-Type", "value": "multipart/form-data; boundary=Boundary123"}
       ],
       "postData": {
         "mimeType": "multipart/form-data",
         "text": "--Boundary123\r\nContent-Disposition: form-data; name=\"account_id\"\r\n\r\n{account_id}\r\n--Boundary123\r\nContent-Disposition: form-data; name=\"text\"\r\n\r\n{post_text}\r\n--Boundary123--\r\n"
       }
     }
   })

   Successful response: {"object":"PostCreated","post_id":"..."}

If Unipile is not connected:

• Option A: Copy the post text for manual pasting
• Option B: Save as a draft file for later publishing

After publishing, suggest:

• "Share the post URL in your course group for cross-engagement"
• "Set a reminder to respond to comments in 2-4 hours"

Voice Guidelines

When SOUL.md exists, use it as the primary voice reference. When it doesn't, use these defaults:

• Professional but human (not corporate)
• Specific over vague (numbers, names, concrete examples)
• Personal perspective over generic advice
• Conversational, not academic
• No buzzwords: "leverage," "synergy," "game-changer," "unlock"