Skip to content

SKILL.md

Latest commit

 

History

History
156 lines (112 loc) · 9.56 KB

SKILL.md

File metadata and controls

156 lines (112 loc) · 9.56 KB
name linkedin-algorithm-skill
description Score and optimize LinkedIn posts using reverse-engineered algorithm signals (2026). Built from 5 peer-reviewed LinkedIn Engineering papers (360Brew, LiGR, Feed SR), LinkedIn VP/Director statements, and large-scale datasets (van der Blom 1.8M posts, AuthoredUp 3M+ posts, Buffer 52M posts). Use this skill whenever the user wants to write, optimize, score, or improve any LinkedIn content including posts, carousels, videos, newsletters, or articles. Also use when the user mentions LinkedIn algorithm, LinkedIn engagement, LinkedIn reach, post optimization, carousel strategy, LinkedIn hook, dwell time, golden hour, or wants to understand why a LinkedIn post performed well or poorly. This skill focuses specifically on algorithm-level optimization and scoring (not general content creation) -- use it instead of general social media skills when the user wants to maximize reach, engagement, or algorithmic performance on LinkedIn.

LinkedIn Algorithm Optimizer

Optimize LinkedIn content using reverse-engineered ranking signals from LinkedIn's production feed system.

Source Material

This skill is built from three rounds of primary-source research:

  1. 5 peer-reviewed arXiv papers on LinkedIn's feed ranking system (360Brew, LiGR, Feed SR)
  2. Ranking signals reference: references/ranking-signals.md (format data, timing, engagement hierarchy)
  3. Technical architecture reference: references/technical-architecture.md (feed pipeline, 360Brew, LiGR)

Read references/ranking-signals.md for the full engagement hierarchy and format performance data.

Confidence Framework

LinkedIn's algorithm is NOT open-source (unlike X/Twitter). Claims in this skill are tagged:

  • CONFIRMED: LinkedIn official statement or engineering paper
  • LARGE-SCALE DATA: 10K+ post study with named methodology
  • PRACTITIONER ESTIMATE: Expert-derived, not independently verified

When scoring or advising, distinguish what we know from what we infer. Never present practitioner estimates as confirmed facts.

How LinkedIn's Feed Actually Works

LinkedIn's feed is a 4-stage neural pipeline (NOT the simplified "show to 5% then expand"):

  1. Retrieval (<50ms): LLM-based dual encoder narrows hundreds of millions of candidates to ~2,000 [PEER-REVIEWED: arXiv 2510.14223]
  2. Ranking: Sequential transformer (Feed SR) scores candidates using your last 1,000 interactions, predicting both passive (dwell) and active (like, comment) engagement [PEER-REVIEWED: arXiv 2602.12354]
  3. Re-ranking: Attention-based diversity ensures you don't see the same author/topic repeatedly [PEER-REVIEWED: arXiv 2502.03417]
  4. Serving: Trust classifiers, language matching, deduplication filters applied

The system refreshes embeddings every 30 minutes. Early engagement within the first 60-90 minutes matters because the model re-evaluates content at each refresh cycle [CONFIRMED in papers + PRACTITIONER CONSENSUS].

Step 1: Determine Format

Content formats ranked by engagement (multiple large-scale datasets agree):

  1. PDF carousel/document (highest engagement across all datasets) [LARGE-SCALE DATA]
  2. Poll (high engagement but LinkedIn deprioritizing in 2026) [PRACTITIONER ESTIMATE]
  3. Native video (vertical, <60s, with captions) [LARGE-SCALE DATA]
  4. Multi-image post (faces increase engagement 38%) [LARGE-SCALE DATA]
  5. Text-only post (800-1,000 characters optimal) [LARGE-SCALE DATA]
  6. Article / newsletter (lower feed reach but longer lifespan + SEO value)
  7. Link post (underperforms all native formats) [LARGE-SCALE DATA]

Decision tree:

  • How-to, framework, checklist? -> Carousel (8-12 slides, 1080x1350, PDF)
  • Thought leadership, opinion? -> Text post (800-1,000 chars, strong hook)
  • Demo, behind-the-scenes? -> Native video (<60s, vertical, captions)
  • Quick engagement? -> Poll (use sparingly)
  • Deep analysis, 1000+ words? -> Newsletter (subscribers get triple notification)
  • Linking external content? -> Native content first, link in first comment only

Step 2: Analyze the Draft

Evaluate against confirmed ranking signals:

Will this generate dwell time? [CONFIRMED signal]

  • Long enough to require 45+ seconds of reading?
  • Strong hook in first 2 lines (triggers "see more" click)?
  • Readable structure (short paragraphs, line breaks, 6th-8th grade level)?
  • Mobile-friendly formatting (70%+ users are on mobile)?

Will this generate quality engagement? [CONFIRMED]

  • Invites meaningful comments (15+ words valued more than "great post")?
  • Ask a genuine question or share a debatable opinion?
  • Will you reply to comments within the first hour?

Will this trigger saves/shares? [LARGE-SCALE DATA: 5x reach correlation]

  • Contains a framework, checklist, or reference-worthy content?
  • Would someone bookmark this for later?

Will this trigger negative signals? [CONFIRMED]

  • Engagement bait ("Like if you agree") -> detected and penalized
  • Pod-like engagement patterns -> actively detected
  • External link in post body -> underperforms (behavioral, not explicit penalty)

Step 3: Optimize for the Algorithm

The Golden Hour: Content is re-evaluated every ~30 minutes. Front-load engagement:

  • Post when your audience is active (Tue-Thu, 5-6 PM local; Wednesday best) [LARGE-SCALE DATA: Buffer 52M posts]
  • Reply to every comment in the first hour (+30% engagement boost) [LARGE-SCALE DATA: Buffer 2M posts, VERIFIED]
  • Have 2-3 colleagues ready to leave meaningful (15+ word) comments early

Hook mechanics: The "see more" click is a confirmed engagement signal. First 2 lines must stop the scroll:

  • Contrarian: "Everyone says [advice]. Here's why that's wrong."
  • Data surprise: "[Specific number] changed how I think about [topic]."
  • Personal failure: "I lost [thing] because I [mistake]."
  • Bold claim: "The best [professionals] do this one thing differently."

Hashtags: 3-5 maximum. They are metadata only, not discovery. 6+ hashtags correlates with significant reach drop [CONFIRMED: LinkedIn VP of Product stated hashtags don't affect distribution; LARGE-SCALE DATA supports 3-5 optimal].

Links: Don't put links in the post body. Link posts underperform all native formats. If you must share a link, put it in the first comment. LinkedIn's Sr. Director of Product denied an intentional penalty, but the behavioral effect is real [CONFIRMED denial + LARGE-SCALE DATA showing underperformance].

AI content: LinkedIn uses C2PA for images/video (opt-in metadata) but has NO confirmed text-based AI detection system. However, AI-generated text may underperform because it lacks the specificity and personal experience that drives dwell time and comments [PRACTITIONER ESTIMATE].

Step 4: Score the Post

Rate each draft on this engagement-signal-aligned scale:

Category Score 1-10 Weight Based on
Dwell potential (will people stop and read?) _ 30% CONFIRMED
Conversation potential (will people comment meaningfully?) _ 25% CONFIRMED
Save/share potential (bookmark-worthy?) _ 20% LARGE-SCALE DATA
Format optimization (right format for this content?) _ 15% LARGE-SCALE DATA
Negative risk (how SAFE is this post? 10=no risk, 1=high risk) _ 10% CONFIRMED

Weighted Score = (Dwell * 0.3) + (Conversation * 0.25) + (Save * 0.2) + (Format * 0.15) + (Safety * 0.1)

Note: Safety scores positively. A safe post (no bait, no pods, no links) scores 9-10. A risky post (engagement bait, external links) scores 1-3.

Score Rating Action
8.0+ Elite Post with confidence
7.0-7.9 Strong Post, minor tweaks optional
6.0-6.9 Average Review critical signals before posting
Below 6.0 Weak Rewrite, focus on dwell + conversation

Step 5: Rewrite Suggestions

When suggesting rewrites, explain which signal you're targeting:

Example:

  • Original: "We just launched a new feature for our product."
  • Problem: Low dwell (nothing to read), low conversation (no reason to respond), low save (nothing to reference)
  • Rewrite: "We spent 4 months building a feature nobody asked for. Then it became our most-used feature overnight. Here's what we learned about building for needs people can't articulate yet."
  • Why: Opens curiosity gap (dwell), invites debate on product strategy (conversation), contains a lesson worth saving (save)

Anti-patterns

These behaviors reduce reach based on confirmed signals:

  1. Engagement bait: "Like if you agree," "Comment YES" -> LinkedIn detects and penalizes [CONFIRMED]
  2. Engagement pods: Same people engaging immediately every time -> actively detected [CONFIRMED: 97% detection rate claimed, single source]
  3. Over-posting: More than 5x/week shows diminishing returns [LARGE-SCALE DATA]
  4. Editing within 10 minutes: May reset evaluation cycle [PRACTITIONER ESTIMATE, unverified]
  5. Excessive hashtags: 6+ hashtags correlates with reach drop [LARGE-SCALE DATA]
  6. External links in body: Behavioral underperformance across all datasets [LARGE-SCALE DATA]
  7. Non-compliant automation: Browser automation and scraping tools are banned. API-compliant scheduling tools (Buffer, Hootsuite) are safe [CONFIRMED]

Output Format

When optimizing a post, provide:

  1. Format recommendation (carousel / text / video / newsletter + why)
  2. Signal analysis (which ranking signals this will trigger, which it won't, with confidence tags)
  3. Score (weighted 1-10 scale above)
  4. Specific rewrites (with signal rationale for each change)
  5. Posting strategy (timing, first-hour plan, comment reply strategy)