--- name: deep-research-post-engine description: > Researches topics deeply across multiple sources, then writes data-backed LinkedIn posts that match your writing voice. Use when asked to "write a LinkedIn post", "create social media content", "help me post about", "research and write about", "draft a post on", or "write content for LinkedIn". Also triggers on "help me grow on LinkedIn", "LinkedIn content strategy", or "write a thought leadership post". Produces research-first content with real sources — not generic AI writing. license: MIT metadata: author: tomer-ezri version: "1.0.0" category: content-creation tags: [linkedin, content, research, social-media, writing] --- # Deep Research Post Engine Research topics deeply, then write data-backed LinkedIn posts that sound like you — not like AI. --- ## Onboarding ### First-Time Detection Ask: **"Have you used Deep Research Post Engine before?"** If new user, show: ``` Welcome to Deep Research Post Engine! This skill researches your topic across 5-8 real articles, extracts hard data and contrarian angles, then writes a LinkedIn post in YOUR voice — not generic AI writing. What makes it different: - Real research with real sources (not hallucinated stats) - Anti-slop filter catches "navigate", "landscape", "game-changer" etc. - Your voice profile ensures posts sound like YOU wrote them - 3 angle options so you pick the framing Let's get you set up. ``` ### Startup Sequence (Tool Detection + Context Questions) At the start of every run, do these **in parallel**: **A. Background Tool Detection (run silently while asking questions):** Search for these tools/MCPs without blocking the user: | Tool | What it does | How to detect | |------|-------------|---------------| | read-website-fast MCP | Reads full articles, not just snippets | Check if `mcp__read-website-fast__read_website` tool is available | | Apify MCP | Scrapes LinkedIn profiles, paywalled articles, social media | Check if `mcp__apify__*` tools are available | **B. Context Questions / First-Time Setup (ask immediately, don't wait for tool detection):** 1. Ask user for **2-3 of their best-performing LinkedIn posts** (paste full text) 2. Ask for their **niche** (e.g., "AI in healthcare", "B2B SaaS growth") 3. Ask for their **target audience** (e.g., "CTOs at mid-market companies") 4. Ask for their **language preference** (English, Hebrew, bilingual, etc.) 5. Ask (optional): **Do you have an existing content calendar or marketing strategy doc?** If yes, read it and align post topics with the existing plan. **Skip any question** where the answer is already known from Claude's memory, prior conversations, or information the user provided in their initial message. If ALL answers are already known, confirm them briefly and proceed. **Wait for the user to answer all remaining questions before starting any analysis.** ### Tool Detection Results (after background check completes) Once background detection finishes AND user has answered context questions: 1. **If all recommended tools are available** → Announce tools and proceed. 2. **If some tools are missing** → Tell the user which tools are missing, what they enable, and ask: "Would you like help installing the missing tools? Or should I continue without them?" - **If yes** → Provide install commands and guide them through setup: - read-website-fast: `claude mcp add read-website-fast -s user -- npx -y @just-every/mcp-read-website-fast` - Apify MCP: 1. Get a free API token at https://console.apify.com/account/integrations 2. Install: `claude mcp add apify -e APIFY_TOKEN=your_token -- npx -y @apify/actors-mcp-server` - **If no** → Continue, but announce: "Proceeding without [tool]. Research depth may be reduced without full article reading, and I won't be able to scrape your LinkedIn profile for voice matching." ### Profile Scraping (for voice profile building) When building or updating the user's voice profile, scrape their LinkedIn profile to understand their writing style, audience, and brand: **Scraping fallback chain (pick first available):** 1. **Apify MCP** (best) — Use Apify's LinkedIn scraper to get the user's profile, recent posts, and engagement data. Ask the user: "I can scrape your LinkedIn profile to better understand your voice and audience. Would you like me to do that?" If user declines, skip to next fallback. 2. **read-website-fast MCP** — Fetch the user's LinkedIn profile URL directly. May get limited data due to LinkedIn's restrictions. 3. **WebFetch (built-in)** — Basic web fetch of the profile URL. Will likely get minimal data. 4. **Ask user to paste** — "I couldn't access your LinkedIn profile automatically. Please paste the text from 2-3 of your recent posts so I can build your voice profile." After scraping, extract: writing patterns, typical post length, formatting style, emoji usage, hook patterns, CTA patterns, topic themes, engagement patterns. Use this data to enrich the voice profile saved to `references/voice-profile-template.md`. ### First-Time Voice Profile Build After context questions are answered and profile scraping is complete (or skipped): 1. Build a voice profile from their posts and scraped data 2. Save to `references/voice-profile-template.md` Voice profile captures: sentence length, punctuation habits, emoji usage, vocabulary level, signature phrases, formatting preferences (bullets vs paragraphs, line breaks, bold usage), and tone (direct, conversational, authoritative, vulnerable). --- ## Tool Detection + Fallback Chain At the start of **every run**, detect available tools and announce them: ### Web Reading (pick first available) 1. **read-website-fast MCP** — Best. Reads full article content, not just snippets. 2. **WebFetch** (built-in) — Adequate. Works for most public articles. 3. **WebSearch** (built-in) — Minimum. Search snippets only, no full article reads. ### Optional Enhancements - **Apify RAG MCP** — Unlocks paywalled articles and social media post research. - **Agent Teams** — Enables parallel research (4 agents at once instead of sequential). ### Announce to User If read-website-fast is available: ``` Tools: read-website-fast (deep article reading), WebSearch. Ready for full research. ``` If only built-in tools: ``` Tools: WebFetch + WebSearch (basic). Install read-website-fast for deeper research: claude mcp add read-website-fast -s user -- npx -y @just-every/mcp-read-website-fast ``` --- ## Agent Teams Roles When Agent Teams is enabled, spin up parallel research agents: | Role | Agent Name | Job | |------|-----------|-----| | Orchestrator | main | Coordinates research, assembles final post, talks to user | | Trend Scanner | research-trends | Searches for latest news and developments on the topic | | Data Hunter | research-data | Finds statistics, studies, hard numbers with sources | | Contrarian Finder | research-contrarian | Finds what most people get wrong, opposing viewpoints | | Audience Connector | research-audience | Connects the topic to the user's specific niche and audience | **When Agent Teams is NOT available:** Run all 4 research roles sequentially in the main session. Same research depth, just slower. --- ## Research-to-Post Pipeline ### Phase 1: TOPIC RESEARCH (5-10 minutes) 1. User provides topic + optional angle 2. Run **8-12 web searches** across different angles: - "[topic] latest news 2026" - "[topic] statistics data" - "[topic] controversy debate" - "[topic] expert opinion" - "[topic] [user's niche] impact" - "[topic] what most people get wrong" - "[topic] case study results" - "[topic] future predictions" 3. Read the **top 5-8 articles** fully (use best available reading tool) 4. Extract from each article: - Hard data points (numbers, percentages, study results) - Direct quotes from named experts - Contrarian or surprising findings - Connection points to user's niche 5. Compile an internal research brief (do NOT show raw brief to user) ### Phase 2: ANGLE SELECTION Present **3 angles** to the user: **Angle 1 — Contrarian:** "The data says the opposite of what everyone thinks..." Best when: research uncovered surprising findings that challenge conventional wisdom. **Angle 2 — Analysis:** "Here's what [development] means for [your audience]..." Best when: there is a recent event or trend with non-obvious implications. **Angle 3 — Personal Experiment:** "I tested [topic] — here's what happened..." Best when: user has hands-on experience or the topic lends itself to a "try it yourself" frame. Include a **recommendation** based on shareability and data strength. User picks an angle or suggests their own. ### Phase 3: POST WRITING Apply the voice profile from `references/voice-profile-template.md`, then write using this structure optimized for LinkedIn algorithm 2026: **Hook** (first 2 lines) — Stop the scroll. Use a bold claim, surprising stat, or pattern interrupt. These lines appear before "...see more" so they must earn the click. **Setup** (2-3 lines) — Frame the problem or context. Why should the reader care right now? **Meat** (bulk of post) — The insight, data, or story. Embed real data points with attribution. Use short paragraphs (1-2 sentences each) for mobile readability. **Pivot** (1-2 lines) — Your personal take or lesson. This is where the user's voice matters most. **CTA** (final line) — A question that invites comments, or "Link in comments" if sharing a resource. **Rules:** - Embed data points with real, named sources inline - Optimize for dwell time (readers should slow down and read, not skim) - Stay under **3,000 characters** (LinkedIn's sweet spot) - Write a separate **"first comment"** containing links, additional context, or thread continuation - First comment must include at least 2 source links with titles (not just raw URLs) **Visual Rhythm (non-negotiable):** Posts live or die by how they look before anyone reads a word. Short lines. Then a longer sentence or two that explains the idea, adds context, and keeps people reading because it feels natural. Then short again. Vary block sizes deliberately: short → long → medium → short → one word. Never uniform. Every post must pass the **squint test** (see Phase 4, check #3) before it leaves the skill. ### Phase 4: QUALITY CHECK (Anti-Slop Filter) Run every post through these checks before presenting to user: 1. **AI slop scan** — Flag and remove words/phrases from `references/anti-slop-patterns.md`. Common offenders: "navigate", "landscape", "game-changer", "paradigm shift", "in today's world", "let's dive in", "at the end of the day", "it's important to note". 2. **Source verification** — Every data point must have a real, checkable source. No "studies show" without naming the study. No "experts say" without naming the expert. 3. **Squint test (visual rhythm)** — Zoom out or squint at the formatted post until the words are barely legible. Look only at the shape of the text blocks. They must vary in height: short, long, medium, short, one word. If all blocks look the same size, the post will feel like a wall — rewrite until the rhythm is visible. Tom Orbach (500+ posts, Forbes 30 Under 30): "I only hit Post when it's beautiful." This check is non-negotiable. 4. **Smart Brevity pass** — Read every sentence and ask: does this word earn its place? Cut every sentence that restates what the previous one said. Cut every adjective that doesn't change the meaning. The goal is the minimum words needed to land the idea — no more. 5. **Voice match** — Compare sentence length, tone, and vocabulary against the voice profile. Flag any sections that drift toward generic LinkedIn-speak. 6. **Expert test** — Ask: "Would someone with 10 years of experience in this field write this, or would they cringe?" Remove anything that fails this test. 7. **Deliver 2 versions:** - **Standard** — Polished, professional, ready to post - **Punchier** — Shorter sentences, bolder claims, more personality 8. **Brand guardrails check** — If the voice profile has brand guardrails, verify the post doesn't violate any of them. Check product naming, positioning, and framing against the guardrails list. ### Phase 5: POSTING GUIDE After the user picks their preferred version, provide: - **Best time to post** — Based on their audience's likely timezone and LinkedIn engagement data - **Image/visual suggestion** — What type of visual would increase engagement (chart, photo, carousel, none) - **Hashtags** — 3-5 relevant hashtags (not spammy, not too broad) - **First-60-minutes plan** — What to do after posting: reply to every comment, engage on 5 other posts in your niche, share to relevant groups - **"Inspired by" amplification** — If the post draws on someone's idea, research, or work, add "Inspired by @[their handle]" at the end. Tag them. People who feel credited as inspiration share. This is one of the highest-ROI distribution moves on LinkedIn — a single share from the right person can 10x reach. - **Follow-up post topic** — A natural next post that builds on this one's momentum ### Phase 5.5: USER REVIEW (before visual report) After completing the post (Phase 4 quality check) and posting guide (Phase 5), present both post versions and the posting guide to the user and ask: **"Here are your two post versions and the posting guide. Would you like to change anything, or should I generate the visual report?"** Wait for the user's response: - **If user requests changes** → Apply the changes (rewrite sections, adjust tone, add/remove data points, etc.), re-run quality checks, and ask again. - **If user approves** → Proceed to Phase 6 (visual report generation). - **If user says nothing specific** → Treat as approval and proceed to Phase 6. ### Phase 6: VISUAL REPORT (browser preview) Generate a self-contained HTML report and open it in the user's browser. 1. Create a JSON object with all post data (see schema below) 2. Read the HTML template from `assets/report-ui.html` 3. Replace `{{REPORT_DATA}}` with the JSON string 4. Save to `./deep-research-report-[topic-slug]-[date].html` 5. Open in browser: `open` (macOS), `xdg-open` (Linux), `start` (Windows) 6. Tell the user: "Visual report saved to [path] and opened in your browser." JSON schema for the report: ```json { "meta": { "topic": "string", "niche": "string", "audience": "string", "date": "YYYY-MM-DD", "toolsUsed": ["string"] }, "research": { "articlesRead": 0, "dataPoints": [ { "fact": "string", "source": "string", "url": "string" } ], "searchQueries": ["string"] }, "angles": [ { "name": "string", "type": "contrarian|analysis|experiment", "description": "string", "recommended": true } ], "selectedAngle": "string", "posts": { "standard": { "text": "string", "charCount": 0 }, "punchier": { "text": "string", "charCount": 0 } }, "firstComment": "string", "quality": { "slopScore": 0, "slopFlags": ["string"], "voiceMatch": 0, "expertTestPass": true }, "postingGuide": { "bestTime": "string", "visual": "string", "hashtags": ["string"], "first60min": "string", "followUpTopic": "string" } } ``` --- ## Output Format Final output includes these sections (copy-paste ready): ``` POST ──────────────────────────────────────── [Full post text, ready to paste into LinkedIn] FIRST COMMENT ──────────────────────────────────────── [Links, sources, additional context] STATS ──────────────────────────────────────── Characters: [count] / 3,000 Data points: [count] (all sourced) AI slop score: [X/10] (0 = fully human-sounding, 10 = AI-obvious) Voice match: [percentage] alignment with your profile POSTING GUIDE ──────────────────────────────────────── Best time: [day + time range] Visual: [suggestion] Hashtags: [3-5 hashtags] First 60 min: [engagement plan] Next post: [follow-up topic idea] ``` See `assets/post-template.md` for the full template with examples. --- ## Troubleshooting **"Skill didn't research anything"** Check if web search tools are available. Run tool detection again. If no web tools work, paste article URLs manually and the skill will work from those. **"Post sounds generic"** Voice profile may be missing or outdated. Re-run first-time setup with fresh LinkedIn post examples. The more posts you provide, the better the voice match. **"No data points found"** Topic may be too niche or too new for published data. Try broader search terms, or pivot to a "personal experiment" angle that relies on your own experience instead of external data. **"Post is too long"** Reduce from 5 key insights to 3. Cut the setup section to 1 line. Remove the weakest data point. Ask the skill to rewrite with a stricter character limit. **"Sources seem wrong"** Report which data point looks off. The skill will re-search that specific claim and either find a valid source or remove it. Never publish unverified stats.