Contrarian Corpus
A Claude MCP server

Teach Claude to argue
like an activist investor.

Built from 34,166 slides across 848 real decks — from Pershing Square, Muddy Waters, Elliott, Starboard, Hindenburg, Spruce Point and 32 more — spanning 2005 to 2026. Not advice, not opinion — just the rhetorical grammar of the investors who actually move markets, served as an MCP your editor queries live.

Why it exists

Most AI sounds nothing like a real investor memo.

Ask any LLM to write an "activist thesis" and you get a polite essay: balanced, hedged, forgettable. Ask a human who has run a campaign and you get something sharper — a diagnosis, a named villain, a number you can't unsee, and a single slide where the trap snaps shut.

The gap isn't intelligence. It's grammar. Pros build arguments with a small, repeating set of moves — peer-gap charts, sum-of-parts reveals, CEO-quote contradictions. We codified that grammar from evidence, not instinct.

What we found

Activists don't argue — they expose.

Every slide in the corpus was classified by narrative function. The pattern that emerges is startlingly consistent across firms and decades.

expose contradiction 25.1% of all slides. The single most common narrative function — management says X, reality shows Y.
decks with a named villain 70.0% The target is almost always someone, not just a stock — a specific CEO, chair or auditor.
decks citing CEO's own words 52.6% Own-goals beat external critique. Pulled earnings-call and letter quotes are the sharpest weapon.
decks with sum-of-parts 26.9% When value is hiding, the move is to decompose — break the target into parts and price each one.

Explore all distributions →  ·  Browse 37,061 extracted phrases →

What you can ask

Plain-language tasks Claude handles well.

The MCP is active whenever your prompt looks contrarian-research-shaped. Claude pulls only the files it needs from the live corpus — no pre-loading, no tarball to sync.

Evaluate a company

"Use the contrarian skill to find the strongest short thesis against Tesla today — three reasons, peer-gap evidence, named villains."

Review a PDF

"Read this earnings-call transcript and use the contrarian patterns to pull every CEO-quote contradiction worth weaponising."

Build a campaign deck

"Draft an activist deck for [target]. Initial thesis phase. Pick 3 patterns and a sum-of-parts valuation frame."

Stress-test a memo

"Here's my long thesis on [X]. Argue the other side with the sharpest patterns in the corpus. Don't be polite."

Letter to the board

"Write a letter to the board of [company] demanding a breakup. Use the governance thesis frame and the primary-demands structure."

Slide architecture

"Design an 8-slide deck for a proxy fight at [target]. Use the SCQA backbone and include one peer-gap chart."

How it actually works

On-demand, not pre-loaded. Live, not cached.

Claude sees the corpus as a browsable filesystem of patterns. When your prompt is contrarian-research-shaped, it reaches in and pulls only the files it needs for that specific answer — nothing more, nothing less. We ship an update, you feel it on the next call.

Decision-tree taxonomy. The corpus is organised by the choices a trained analyst makes in order: pick the output form, locate the campaign phase, classify the thesis, pick rhetorical patterns, anchor the valuation, design the slide architecture, carry the narrative. Each node is prescriptive — "use median, never mean", "5–7 peers, not 3", "one KPI per chart" — not advisory.

Named exemplars. Every pattern cites real decks with dates — Canadian Pacific 2012, Herbalife 2012, Arconic 2017, Luckin 2020, Phillips 66 2025. Claude doesn't imagine what a good peer-gap chart looks like; it cites the ones that worked.

Live at the endpoint. The MCP exposes search_patterns, list_categories, and get_pattern — plus every file as a readable resource. Claude browses, searches, pulls. New patterns we add today are in your next answer. No re-install.

No scaffolding in the output. The files are written as generic domain knowledge — no meta-talk, no self-references. The output speaks for the work; the reader never sees the machinery.

Why MCP

A live endpoint your editor queries — instead of a tarball you forget to update.

Always current

Every new pattern, every amended example, is live on your next tool call. No re-download, no version-skew between you and us.

Usage signal

Per-token telemetry lets us see which patterns get asked for and which don't — a feedback loop that static files can't give.

Scoped access

Tokens are revocable. If you rotate, we rotate. Email encrypted at rest — only the offline admin key can decrypt it.

Install

Register. Install. Use.

One email, one command, no dependencies. Your token stays tied to that email — regenerate anytime by re-submitting.

1 · Get your token

We store a sealed-box-encrypted copy of your email (only the offline admin key decrypts it) and a SHA-256 hash of the token.

Your token · shown once.

2 · Point Claude at it
npx @anlakstudio/contrarian-install <your-token>

Detects your Claude setup (Code, Desktop or Cursor) and writes the MCP config block. Idempotent — safe to re-run. Restart Claude and the contrarian skill is available.

When to invoke

Built for specific jobs.

  • Building an activist campaign thesis — long with proposed changes.
  • Writing a short-seller report — alleging overvaluation, fraud or broken model.
  • Structuring a contrarian investment memo for internal circulation.
  • Drafting a letter to the board or to shareholders.
  • Designing the slide architecture for any of the above.
  • Critiquing a prevailing consensus with a structured counter-argument.
Who made this

Built by Anlak. Every document is linked to its original source with attribution. For commercial advisory work that applies this corpus to specific client situations, get in touch or email miguel@anlak.es.

Attribution & takedowns · Browse 848 decks · 12 patterns