media-diet

Media Diet

Algorithm intelligence for social feeds. A browser extension reads the feed; Claude reads the algorithm.

Most "algorithm transparency" tools either guess (from outside, with no access) or scrape aggressively (fragile, ToS-hostile). Media Diet does neither. A Chrome MV3 extension runs inside your own authenticated session and extracts the feed via direct DOM access — free, structured, deterministic. Claude Code then synthesizes an editorial report on what the algorithm is doing to you, and generates an AgenticHarness: a JSON tool definition that any AI agent can consume to re-analyze the same profile, forever.

One scan. Infinite re-analysis. Agent-agnostic by design.

The hackathon build

Built for the Anthropic Hackathon (April 2026) by Nathaniel Garelik, Ben Verhaalen, and Aiden Lang.

What's working end-to-end:

• Chrome MV3 extension with content scripts for Instagram, X (Twitter), YouTube, and GitHub
• DOM-based feed extraction with scroll + dedup, producing structured FeedItem[] that persist to disk
• Profile synthesis: feed → AlgorithmProfile via Sonnet 4.6 (metrics: geographic concentration, topical diversity, political lean, creator tier mix, echo index)
• Harness generation: profile → AgenticHarness (JSON tool definitions, OpenAPI-compatible schemas) via Sonnet 4.6
• Six-command Claude Code plugin with Orchestrator (Sonnet 4.6) + Analyst (Haiku 4.5) subagents
• Natural-language router (/mesh <intent>) dispatches across scan, bias, status, tools, and compare
• Durable local storage — every scan, profile, harness, and report is a JSON file under data/

The sample report above is real. Unedited output from /mesh-scan instagram against an actual authenticated Instagram session during development. The only changes are handle and name placeholders; structure, scoring, and voice are exactly what Claude produced.

Demo walkthrough

A five-minute flow that shows the full pipeline:

1. Install — npm install && npm run build:extension, then load extension/dist/ as an unpacked Chrome extension.
2. Scan — navigate to instagram.com in your authenticated browser and click the extension's Scan button. Raw feed items land in data/scans/instagram/{timestamp}.json in seconds.
3. Synthesize — from Claude Code with the Media Diet plugin installed, run /mesh-scan instagram. The Orchestrator (Sonnet) builds an AlgorithmProfile and a matching AgenticHarness.
4. Analyze — /mesh-bias instagram produces the full editorial report (like the sample above). Or ask anything in natural language via /mesh <intent> — the Analyst subagent (Haiku) answers against the cached profile, no rescan needed.
5. Compare — scan X or YouTube, then /mesh-compare to see how different algorithms categorize the same person differently. This is where the agent-agnostic harness story shines: the same JSON tool defs that drive Claude's analysis could drive any other agent's.

Why Claude

Three decisions that shape the architecture:

1. Claude does what only Claude does well. DOM extraction is cheap and deterministic — we don't waste a model on it. Claude's job is synthesis: turning a list of feed items into an editorial verdict with voice, hedged claims, and concrete advice. That's the work models are uniquely good at.
2. Model routing is a first-class design concern. The orchestrator (Sonnet 4.6) plans and synthesizes the profile once. The analyst subagent (Haiku 4.5) answers every follow-up question against the cached profile at ~10× lower cost. One scan → ~$0.02 of Sonnet. Every subsequent question → ~$0.001 of Haiku.
3. Claude Code is the runtime. Ships as a Claude Code plugin: six slash commands, two agents (orchestrator + analyst), and a skill. No separate server, no separate UI, no backend database. Users install it like any other plugin and drive it in natural language.

What a report actually looks like

Real report from one scan, handles and personal details replaced by placeholders. Structure, scoring, and voice are unedited.

---

Media Diet Report · @user

Pulled headlessly from your IG feed right now. No mocks, no API keys, no frontend — this is Claude reading the raw signal.

Who the algorithm is serving you

Tight regional-university bubble dominates the feed: campus_life_a, campus_bar_a, campus_humor_a, campus_humor_b, campus_psych, campus_alum_a, plus the university's institutional account ecosystem. Bias isn't political — it's geographic + institutional. Your algorithm has decided you are a regional-university undergrad first, everything else second.

Secondary shelf: brainrot prank content — prank_a, prank_b, prank_c. Frat-house humor, short-form, high-engagement. This is what's eating your dwell time.

Outliers worth keeping: travel_niche_a (a travel-adjacent nostalgia account, probably seeded by your own travel posts), primevideo / franchise_promo_a, product_ad_a + product_ad_b ads.