Claude Skill – sourced research on how 5 social apps work

What happened

The HTML content is mostly navigation/boilerplate. Let me search for the actual README content in the file. Now I have the detailed content. Let me also search for more specific technical details. I now have enough concrete facts from the repository. Let me write the article.

A GitHub repository called social-app-teardown, published by elchrysaki, set out to answer a question that has frustrated developers and product researchers for years: what do Threads, LinkedIn, Instagram, X, and Reddit actually do under the hood? The project draws on primary sources — 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, and Reddit's original Wilson-score comment-ranking code pulled straight from Reddit's own repository — and cross-checks all of it against direct, live observation of the real running apps. The result is a structured reference for each of the five platforms, with every claim tagged by confidence tier. The repository surfaced on Hacker News under the HN Claude/Codex/Fable tag, though the thread had received 1 point and 0 comments at time of writing.

This piece walks through what the teardown contains, how the research was assembled, what each platform's mechanics reveal about product strategy, who will find the repository most useful, and what builders should verify before relying on it in their own work. It is written for product engineers, UX researchers, and anyone who has been frustrated by recycled "algorithm explainer" content that never cites a source.

How it works

A GitHub user named elchrysaki published a repository called social-app-teardown containing documented teardowns of five major social platforms: Threads, LinkedIn, Instagram, X, and Reddit. The stated goal was to replace guesswork with sourced, citable material on how each platform's feed ranking, features, UX patterns, and brand systems actually work. Each platform folder contains five documents: a feature catalog, an algorithm-and-mindset file, a UX/UI catalog, a brand-and-marketing kit, and a psychology-and-persuasion file. A sixth file, 00-synthesis.md, serves as a cross-platform comparison that the repository readme recommends reading first.

The repository appeared on Hacker News tagged under the HN Claude/Codex/Fable category, which broadly covers AI-assisted research and coding tools. The submission had 1 point and 0 comments at the time of writing. The repository is public and free to read, fork, or build from without restriction. There is no license file listed, which is a detail builders should confirm before incorporating any portion of the text into a commercial product.

The research methodology rests on primary sources rather than secondhand summaries. LinkedIn's March 2026 engineering post describing its LLM-based feed rebuild feeds the LinkedIn algorithm file. Instagram's exact Explore scoring formula comes from Meta's own Transparency Center system cards. X's ranking model is drawn from the two separate releases of X's open-sourced ranking code. Reddit's Wilson-score comment ranking is traced to Reddit's own public repository. Where primary sources were unavailable, the teardown uses direct, authenticated live observation of the real apps — the LinkedIn DM smart-reply chips, for instance, only appeared during an authenticated session and could not be found in any public documentation.

Why it matters

Each claim is tagged by confidence level so a reader can see at a glance whether something is confirmed by official engineering documentation, inferred from observable product behavior, or somewhere in between. The UX/UI catalog documents include extracted design tokens alongside before-and-after visuals of live UI elements. The brand-and-marketing kits document real copywriting patterns, tone parameters, and visual specifications such as exact color values. A dark-patterns audit runs across every documented feature using Harry Brignull's named taxonomy as its framework.

Most widely shared explainers of social platform algorithms are built on a mix of reverse-engineering guesses and recycled observations from other explainers, creating a citation loop with no primary source at the bottom. This project breaks that loop by anchoring every significant claim to a traceable original: a platform's own engineering post, its own transparency page, or its own public code. The March 2026 LinkedIn post is notable because it documents a live architectural shift — the feed now reads a member's last 1,000 or more interactions as a single sequence rather than treating them as independent events, and explicitly excludes demographic attributes from ranking by LinkedIn's own stated design.

Who is affected

Two findings in particular illustrate why sourced teardowns matter for builders. Reddit deliberately randomizes the vote counts users see, a detail documented in Reddit's own codebase, specifically so a vote-manipulation bot can never confirm whether its attack succeeded. Instagram's mobile web wraps every profile-grid link in an Android app-install intent, making post detail accessible to a logged-out desktop visitor but not to a logged-out mobile visitor — a pattern the teardown classifies as a textbook "Sneaking" dark pattern under Brignull's taxonomy.

The repository readme lists three groups as its intended audience: researchers and product people who want a citable, source-tiered account of how these platforms work; marketers and brand strategists who want real design tokens and sourced brand-voice analysis rather than generic style-guide summaries; and product ethicists, UX researchers, and educators who want a rigorous named-framework audit of dark patterns. A fourth group not mentioned explicitly but clearly served is any developer who is building a competing social product and needs a rebuildable UX/UI spec to understand what signals and patterns to replicate, reference, or deliberately avoid.

Threads developers will find the internal documentation detail notable: the platform's own ranking documentation still refers to the product as "the Text Post App" in every signal description, three years after launch, suggesting the internal system has not been renamed in step with the public brand. Instagram developers building mobile-web experiences need to account for the intent:// link wrapping behavior, which has direct consequences for SEO and logged-out user acquisition. Reddit developers and third-party client builders should note that karma functions simultaneously as a trust score, a posting-privilege unlock, and an income-potential gate through Reddit's cash-reward program.

What to watch next

The repository does not list a planned update cadence, so builders who use it as a reference should check whether the primary sources it cites have been updated since the teardown was written. LinkedIn's engineering blog is the most time-sensitive dependency given the March 2026 date of the feed-rebuild post, because LLM-based feed architectures tend to iterate quickly. X's open-sourced ranking code has been released in two separate drops already, so a third release could alter the signal weights the teardown documents. Meta's Transparency Center system cards are the source for Instagram's Explore scoring formula, and any update to those cards would supersede what the repository currently records.

A builder using the psychology-and-persuasion files as a product audit checklist should verify that the Brignull taxonomy version used is current, since the named dark-patterns list has expanded over time. The cross-platform synthesis in 00-synthesis.md draws conclusions that depend on all five platform teardowns being accurate at the same point in time, which becomes harder to guarantee as each platform's products evolve independently. The HN thread had 0 comments at time of writing, so there is no community validation or correction layer yet — the most useful signal to watch for is whether practitioners who specialize in any one of the five platforms publicly confirm or contest the sourcing and claims in that platform's folder.