Mostly skills — with a few MCP servers and hooks — for Claude Cowork and Claude Code. Built for knowledge work — most of it needs no code at all — and honest about what each tool does and doesn't do.
- research-toolkit — Loaded topics. No thumb on the scale — especially yours. Verify a claim, trace who benefits, and turn your own conclusion inside out before you trust it.
- makers-toolkit — Discipline for prompting and co-authoring with AI — built on reasons, not bare orders. Nothing to configure — just skills you invoke.
- vasana-system — Proof of concept. Notice the same regularity in unrelated topics — and, more usefully, in how you and the AI actually work together. Real, modest benefits today; not yet the self-organizing learner it's built toward.
- security-toolkit — Useful friction against obvious mistakes: hooks that block dangerous commands and flag prompt-injection. A basic draft, not a security product.
AI doesn't think for you. At its best, it thinks with you. You hand it a half-formed question, a mess of a document, a thing you can't quite see the shape of — and it hands back something you can work with. That's externalized cognition: thinking pulled out of one head and made into something more than one head can hold. It was never really a solo act anyway. The lone mind sealed in its own skull was always a fiction; AI just makes that plain.
But the same capability can be handed to you or taken from you. Watch where search is going: from a list of sources you pick through to a single answer, generated out of sight, that you're meant to trust. That's externalized cognition at scale — and enclosed. A few firms own the machinery and decide which answer you see. (Enclosure is an old move: fence the commons, then rent people back what they used to use for free.) Cognition can be held in common or walled off the same way, and that fight isn't settled.
This repo doesn't settle it. It's small, and the real work is happening in more places than this one. But here's a handful of tools built the other way: they hand you the means, not the answer — something to drive, not an oracle to trust — and some turn on the machinery itself, asking who owns it, who benefits, and what's being left unsaid. Free to install, local, yours to change.
Claude Cowork — no terminal needed: add the marketplace bogheorghiu/ex-cog-dev
via Customize → Browse plugins, then install the plugins you want.
Claude Code
/plugin marketplace add bogheorghiu/ex-cog-dev
/plugin install research-toolkit@ex-cog-dev
/plugin install makers-toolkit@ex-cog-dev
/plugin install vasana-system@ex-cog-dev
/plugin install security-toolkit@ex-cog-dev
Note
ex-cog-dev is the public, canonical marketplace — add and install from it
directly. The -dev in the name is historical; it stays so anyone who already
added the marketplace keeps working.
Loaded topics. No thumb on the scale — especially yours.
Interrogate what you're sold, and what you'd rather believe: verify a claim, trace who benefits, and turn your own conclusion inside out before you trust it.
- Not sure where to begin?
researchroutes the question for you — by domain, by how deep you need to go, by what sources exist — so you don't have to know the method names to use the methods. - Map power.
cui-bonoasks who gains and who loses, through six lenses — weapons, labor, environment, governance, supply chains, geopolitics.deep-investigation-protocolruns staged source sweeps for when the marketing and the reality diverge.manufactured-consensus-detectionasks whether your sources agree on their own or just echo one origin.source-omission-analysisreads the silences — what no one is saying.dev-job-defense-tiesruns that same buyer-chain logic on a job offer — screen a studio or employer for hidden military or defense ties against a red line you set (kept on your machine, never committed). - Break your own case.
dialectic-spiralbuilds the strongest opposite of your conclusion and runs it at the evidence, four rounds minimum — with a bench of adversarial agents (a falsifier, a critic, a spiral that holds contradictions open instead of resolving them) to do the arguing.text-deconstructionfinds where a text contradicts itself on its own terms.frame-rotationrephrases the problem through another language's grammar to knock you out of English defaults.iterative-verificationstops when the evidence clears the bar, not when you're tired. - Pull the sources.
youtube-researchandsubstack-researchmine practitioner know-how and independent reporting (Substack through a browser scraper, one-time login).video-transcript-extractiongrabs transcripts from captions or local Whisper. - Live data, no keys.
financial-mcpreturns prices, fundamentals, history, and indicators — RSI, MACD, Bollinger — from Yahoo Finance.macro-monitorwatches that data for macro-stress signals.transparency-mcpputs public power on the record: Congress bills, members, and votes (GovTrack), World Bank indicators, and nonprofit 990s (ProPublica).
Discipline for prompting and co-authoring with AI — built on reasons, not bare orders. Nothing to configure, just skills that hold a line.
intrinsic-prompt-design— give the model the reason behind a rule, not just the rule. In practice the reasons part clearly helps, even paired with plain imperatives — anecdotally so far; rigorous testing would be worth doing. Its bolder idea — deliberately leaving space for the model's own agency to fill — is the core experimental proposition.system-pilot— engineering discipline for deterministic systems: define what "done" means, split spec from orchestration from tools, test the seams early, schema first, repair loops that record only verified lessons. Adapted from the Universal CLAUDE.md Protocol.skill-activation-testing— find out whether a skill's description actually makes it fire: a fast blind-router proxy, plus a live firing-counter hook (it ships and works) for the real measurement.
A proof of concept — an early step toward a real ambition. The idea: the same regularity can turn up in places that have nothing to do with each other; an AI not boxed into one field's categories can notice it, and check whether it's a real shared mechanism or just a surface rhyme. Today that's a growing, hand-kept list of specific observations — and they do resurface when the same shape shows up again, which is genuinely useful: catching recurrences across unrelated topics, and (often the bigger win) in the work itself, where you and the AI keep making the same moves or settling into a groove. Each one also accumulates a count of the times it has recurred, and that weight seems to matter in practice — an impression, not yet a measured result. Where it's headed — a system that recognizes and learns these patterns more on its own — is the genuine intention, not a misreading of the name; that fuller version just isn't built yet. Some skills are rougher than others and the bundled agent may be more than it needs — but as an honest first step that already helps, it earns its place.
- The loop.
vasanaflags a candidate mid-work.record-patterncaptures it with structure.find-similarchecks whether it recurs or was a fluke.test-patternchecks whether a saved pattern actually fires and changes anything.pattern-librarybrowses the collection.break-patternandcheck-assumptionsturn the same scrutiny back on your own work. - Beyond the loop.
iterative-loop-engineasks "am I actually done, or did I just stop?" and keeps work cycling until the completion criteria — not fatigue — say done.self-improving-investigationruns research through blind worker agents and dialectic synthesis when the risk of confirming your own bias is high.temporal-shapingdesigns the time-shape of a process — phases, pacing, why something feels off. - Two MCP servers, told straight.
relational-memoryworks: it saves facts, task state, and core principles to disk in layered storage, recalls them by search, and summarizes old entries as they pile up.edge-graphrecords relations as weighted edges that grow heavier each time you cross them, so what recurs rises to the top. Both can flag a relation-type or verb that repeats three times or more. What neither does yet is the bigger ambition — reading those recurrences into higher-order patterns, or letting shared vocabulary consolidate on its own. That part is still a sketch. Today they're a memory system that works in practice, not the pattern engine the design imagines. Storage is local disk; nothing survives an ephemeral cloud session.
A basic draft, not a security product — plain bash hooks that pattern-match on commands and tool output, plus one guided-verification skill. No external tools, no sandbox; anything that doesn't match a pattern sails through. Useful friction against obvious mistakes, not something to rely on for real security.
- Catches prompt-injection in any tool output, MCP results included — loud warning for the blatant stuff, quiet log for the rest, with a path allowlist. It flags and records; it doesn't block.
- Blocks dangerous git before it runs: push to
main/master, force push,reset --hard,--no-verify,branch -D,clean -fd,rm -rf. Blockinggh pr mergeis off by default (merge already passes through branch protection); turn it on interactively with the/pr-merge-guardcommand — which takes effect immediately — or declaratively withEXCOG_BLOCK_PR_MERGE=1. Thepr-merge-guardskill explains it and flips it when you ask ("stop auto-merging," "lock down main"). - Blocks Desktop Commander's process-spawning and config-setting tools, the ones that slip past Claude Code's permission boundaries.
- Answers "am I compromised?" after a supply-chain scare — a poisoned npm or PyPI
package, a trojaned extension — with a guided check of a Windows + WSL2 machine
(
windows-wsl-security-verification): known-bad package versions, persistence, planted SSH keys, shell-rc injection, the Windows AV pass. It verifies, it doesn't harden — and a clean result raises confidence, never proves you're clean.
Each plugin is self-contained — its own plugin.json, skills, agents, hooks, and
optional MCP servers. Plugins reference each other by name, not by path.
The ideas and direction are Bogdan Gheorghiu's; the words are mostly Claude's (Anthropic). A person choosing what's worth saying and an AI finding the words for it is the externalized cognition this repo is about — so we'd rather sign it than pretend otherwise.
MIT