social-app-teardown

How Threads, LinkedIn, Instagram, X, and Reddit actually work — every feature, the real ranking algorithms, and a rebuildable UX/UI spec, all cited.

Most "how the algorithm works" content on the internet is guesswork dressed up as authority. This repo isn't that. It's built from primary sources wherever platforms have published them — engineering blog posts, open-sourced ranking code, official transparency pages — cross-checked against direct, live observation of the real apps, with every claim tagged by how confident it is. Where something genuinely couldn't be verified, that's stated outright instead of papered over with a confident-sounding guess.

What's in here

• Every feature, cataloged and categorized — not "Instagram has a report button" but the exact 9 reasons in the live report flow, in the order they're shown, and what happens after you submit one.
• How the ranking algorithms actually work, sourced from the platforms themselves: LinkedIn's own March 2026 engineering post on its LLM-based feed rebuild, Meta's Transparency Center system cards (including Instagram's exact Explore scoring formula), X's two open-sourced ranking-code releases, Reddit's original Wilson-score comment ranking straight from its own repo, and the actual matrix-factorization math behind X's Community Notes.
• A rebuildable UX/UI spec — not "the button looks nice," but the exact hex color, corner radius, padding, and (where the animation is real CSS) the precise timing function, extracted live via getComputedStyle(), plus a named psychological mechanism (Fitts's Law, the Zeigarnik effect, loss aversion, variable reward schedules...) explaining why each design choice works.
• A brand & marketing kit per platform — design tokens (color/type/spacing) ready to reuse, brand voice and positioning backed by quoted evidence, the acquisition-funnel and lifecycle-marketing mechanics, and real copywriting patterns. Includes findings like: Instagram's famous gradient is a marketing-only asset — the actual product UI runs on a completely different flat blue.
• A cross-platform synthesis — the same report flow, the same ranking philosophy, the same moderation model, compared side by side across all five apps, plus what's genuinely unique to each one and what that reveals about its strategy.

Notable findings

A few things surfaced in the research that are worth the click on their own:

• Instagram's iconic gradient is a marketing-only asset. The real product UI runs on a flat blue (rgb(74,93,249)) that never touches it. — instagram/brand-marketing-kit.md
• X is the only one of the five running a custom in-house typeface (TwitterChirp) — the other four ride the OS system font stack purely for load-speed reasons. — x/brand-marketing-kit.md
• A note only appears on X if people who normally disagree with each other both rate it helpful — the actual bridging-based matrix-factorization math behind Community Notes is public, and it's genuinely elegant. — x/algorithm-and-mindset.md
• Reddit deliberately randomizes the vote counts you see, on purpose — so a vote-manipulation bot can never confirm its own attack worked. — reddit/algorithm-and-mindset.md
• Threads' own internal ranking documentation still calls the product "the Text Post App," never "Threads," in every signal description — three years after launch. — threads/algorithm-and-mindset.md
• Instagram's mobile web silently wraps every profile-grid link in an Android app-install intent — post detail is reachable logged-out on desktop, but not on mobile. — instagram/ux-ui-catalog.md
• LinkedIn's DM smart-reply chips turned up only by actually opening an authenticated session — no amount of logged-out research or documentation surfaced them. — linkedin/feature-catalog.md
• Reddit pays users real cash gated on the same karma number it uses to fight spam — one metric is simultaneously trust score, posting privilege, and income potential. — reddit/brand-marketing-kit.md
• LinkedIn's 2026 feed model reads a member's last 1,000+ interactions as one sequence, not independent events — and explicitly excludes demographic attributes from ranking, by the platform's own stated design. — linkedin/algorithm-and-mindset.md

Who this is for

• Builders — designing a competing or adjacent product (a community platform, a habit-tracking app, a comment system) and want a real reference for what to borrow and why it works, not just what it looks like.
• Researchers and product people — who want a citable, source-tiered account of how these platforms rank content and moderate at scale, without re-deriving it from scratch.
• Designers — who want exact, rebuildable specs for interaction patterns instead of eyeballing screenshots.
• Marketers and brand strategists — who want real design tokens, sourced brand-voice analysis, and growth-funnel mechanics instead of a generic "here's their vibe" writeup.
• Anyone curious why these apps are built the way they are, and why they work on you the way they do.

Start here

00-synthesis.md — the cross-platform comparison. Read this first.

Each feature-catalog.md works through 15 categories: account & identity, content creation, feed & algorithm, per-post actions, social graph, notifications, messaging, discovery, moderation & safety, privacy, monetization, gamification, accessibility, and platform-signature features. Instagram and LinkedIn also include an "authenticated-session addendum" — gaps closed via a real logged-in session, browsed strictly read-only, with all personal/private content excluded.

Reproduce or extend it

This wasn't hand-assembled from vibes — it was produced with two Claude Code skills included in this repo under skills/:

• social-media-app-researcher — the research methodology behind everything above: a 15-category feature taxonomy, a UX/UI cataloging framework with a glossary of real design mechanisms, a brand & marketing framework, and a research playbook (source hierarchy, mobile-first observation, live CSS inspection) — reusable to re-run this research after these apps change, or to point at a different platform entirely.
• platform-content-optimizer — takes the research above (or gathers fresh, cited research if none exists yet) and rewrites a piece of content — a post, caption, thread — to actually reward the specific signals a platform's own documentation says it measures (dwell-time pacing, feed-truncation length, comment-vs-like weighting), rather than generic "post more, use hashtags" folklore. It's deliberately a separate skill from the researcher above — one researches, the other writes; see its SKILL.md for why that split matters.

Methodology and honesty standard

• Source hierarchy: direct observation beats official documentation beats reputable secondary sources beats recalled/unverified information — every non-obvious claim is tagged with its tier.
• Gaps are named, never hidden. Where a platform blocked access (login walls, bot detection, network policy), that's stated explicitly. Reddit's files, for instance, are upfront that reddit.com itself was inaccessible this session — that research leans on Reddit's own Help Center and open-sourced ranking code instead, and says so.
• Every "why it works" claim names a real mechanism, tied to a specific outcome — not a vague adjective.

Research conducted 2026-08-29 through 2026-08-30. Social apps change constantly — treat this as a dated snapshot, not a living document, unless it's been re-run since.

Disclaimer

This is independent research and commentary, not affiliated with, endorsed by, or sponsored by Meta, Threads, Instagram, LinkedIn, X Corp., or Reddit. "Threads," "Instagram," "LinkedIn," "X," "Reddit," and any other product or company names in this repo are trademarks of their respective owners, used here solely to identify and refer to the products being discussed.

All content is derived from information the platforms themselves make public — their live, publicly-accessible web pages, their own published documentation and engineering blog posts, and their own open-sourced code repositories — plus reputable secondary reporting, cross-checked and cited throughout. No private, non-public, or access-controlled systems were accessed. No credentials, security measures, or technical access controls were bypassed. No user data, personal information, or content belonging to third parties is included; where research required an authenticated session, it was the author's own account, browsed read-only, with all personal and private content deliberately excluded from what's published here.

Any quoted material is short, attributed, and used for research, criticism, and commentary. This repository does not reproduce any platform's software, source code beyond short illustrative excerpts of code the platforms themselves have published under an open-source license, logos, or other copyrighted creative assets.

This snapshot reflects each platform's behavior as observed on the dates noted above; platforms change constantly, and nothing here should be relied on as current, complete, or authoritative without independent verification. Nothing in this repository is legal, financial, or professional advice.

If you are a rights holder and believe something here should be corrected or removed, please open an issue — I'll respond promptly.

License

MIT — see LICENSE. Use it, fork it, cite it.