Open it as a Claude Code workspace

Every morning, a Claude agent wakes up and receives a fresh creative challenge — maybe limnology (freshwater lake science), maybe coopering (barrel-making), maybe Mars terrain analysis crossed with EVA procedure planning. Each day is a new domain to explore, a new puzzle to solve. The agent builds a full, professional-grade workspace for whatever the entropy engine surfaces, then pushes it here.

This repo is the result. It grows by one workspace daily, completely autonomously. No human in the loop. Just an AI, an entropy engine, and an ever-expanding collection of workspaces for domains you didn't know you needed.

Built by DaxxSec & Claude (Anthropic)

---

The Problem

AI has an entropy problem. Ask it to "pick something creative" a hundred times and you'll get the same handful of safe, predictable ideas. It's the creative equivalent of a random number generator with a bad seed — low entropy, repetitive output.

Cognitropy (cognition + entropy) is our answer.

A note on the terminology: Yes, we made up a word. Two, actually. Cognitropy = cognition + entropy — the injection of unpredictability into AI creative processes. Cognitropic = the adjective form, describing structures that direct that entropy toward cognition (following the Greek -tropos, "turning toward" — the same root behind phototropic and psychotropic). Is it a real academic term? No. Does it describe a real pattern that didn't have a name? We think so. A cognitropic hash structure is a specific thing: multiple salted cryptographic hashes of a shared seed, reduced via modulo into independent selection indices across distinct categorical pools. That's a mouthful, so we just say "cognitropic." You're welcome.

---

The Cognitropy Engine

How do you make an AI genuinely unpredictable without calling any external APIs? You hash the date.

The engine (cognitropy.py) takes today's date (e.g. 2026-03-27) and runs it through five separate SHA-256 hashes, each with a different salt. Each hash produces a massive 256-bit integer — essentially a huge, chaotic number derived from a simple date string. That number gets reduced via modulo (%) to an index into the relevant pool:

$ ┌────────────────────────────────────────────────────────────────────┐ │ \text{COGNITROPY} \text{ENGINE} │ │ │ │ \text{Step} 1: \text{Hash} \text{the} \text{date} \text{with} \text{different} \text{salts} │ │ │ │ \text{sha256}("2026-03-27") → \text{huge} \text{int} → % 363 \text{domains} │ │ \text{sha256}("2026-03-27" + "\text{secondary}")→ \text{huge} \text{int} → % 363 \text{domains} │ │ \text{sha256}("2026-03-27" + "\text{technique}")→ \text{huge} \text{int} → % 30 \text{methods} │ │ \text{sha256}("2026-03-27" + "\text{spark}") → \text{huge} \text{int} → % 5 \text{templates} │ │ \text{sha256}("2026-03-27" + "\text{crossover}")→ \text{huge} \text{int} → % 10 → <3? │ │ │ │ \text{Step} 2: \text{Assemble} \text{the} \text{assignment} │ │ │ │ \text{Primary} \text{Domain} ──────── "\text{limnology}" │ │ \text{Technique} \text{Modifier} ──── "\text{with} \text{safety} \text{protocol} \text{enforcement}" │ │ \text{Crossover} \text{Check} ─────── 7 (≥3, \text{so} \text{no} \text{crossover} \text{today}) │ │ │ │ \text{Step} 3: \text{Output} │ │ │ │ "\text{Build} \text{a} \text{workspace} \text{for} \text{LIMNOLOGY} │ │ \text{with} \text{safety} \text{protocol} \text{enforcement}" │ │ │ │ \text{On} \text{a} \text{crossover} \text{day} (\text{hash} % 10 < 3, ~30% \text{chance}): │ │ "\text{Fuse} \text{LIMNOLOGY} \times \text{CAVE} \text{DIVING} \text{using} \text{techniques} │ │ \text{from} \text{both} \text{domains}" │ └────────────────────────────────────────────────────────────────────┘ $

Why this works: SHA-256 is a cryptographic hash — even a one-day difference in the input date produces a completely unrelated output number. The selections look random but are fully deterministic: run it twice on the same date, get the same result every time. No external APIs, no randomness source needed — just math.

The domain pool spans 363 wildly diverse fields across 22 categories — volcanology, watchmaking, competitive barbecue judging, Mars terrain analysis, coopering, and 358 more. Combined with 30 technique modifiers, 5 crossover spark templates, and the constraint that crossover domains must come from different categories, that's 18,863,790 unique possible outcomes. The creative constraint is the point. Each day brings an unexpected domain, and the agent rises to meet it.

Try it yourself:

python3 cognitropy.py              # Today's assignment
python3 cognitropy.py 2026-04-05   # Check any date

Sample schedule (seeded by date, every day is a surprise):

Date	Domain	Category	Type
Mar 26	Theme Park Queue Optimization × Vertical Farming	Unusual & Niche	Crossover
Mar 27	EVA Procedure Planning	Space & Aviation	Standard
Mar 28	Permaculture Design	Food & Agriculture	Standard
Mar 29	Driveline Vibration Analysis	Automotive & Engine	Standard
Mar 30	Satellite Communication Protocols	RF/SDR/Signals	Standard
Mar 31	Hydraulic Engineering Fluid Dynamics	Engineering & Technical	Standard
Apr 05	Falconry Bird Training × Security Log Analysis	Outdoor & Adventure	Crossover
Apr 09	Film Restoration × Heraldry	Arts & Creative	Crossover
Apr 11	Brake System Failure Analysis × Dendrochronology	Automotive & Engine	Crossover

The diversity is the point. A workspace for mushroom foraging uses the same structured methodology as one for malware analysis — triage, evidence collection, documentation, reporting. The patterns transfer. The domains are just the fun part.

---

How the Daily Build Works

  ┌──────────┐   ┌──────────────┐   ┌──────────────┐   ┌──────────┐   ┌───────────┐
  │ COGNITROPY│──→│ CLAUDE AGENT │──→│ SECRETS SCAN │──→│ GIT PUSH │──→│ DASHBOARD │
  │  assigns  │   │   builds     │   │  validates   │   │  + stats │   │  + README  │
  │  domain   │   │  workspace   │   │  no leaks    │   │  update  │   │  refresh   │
  └──────────┘   └──────────────┘   └──────────────┘   └──────────┘   └───────────┘
       │                │                   │                 │               │
  5 salted hashes  CLAUDE.md          grep for keys     README badges  Static HTML
  of today's date  .claude/commands/  .pem, .env, .key  + WORKSPACES   regenerated
  → domain+method  context/concepts   API tokens        .md refresh    from engine
  → crossover?     workflows + refs   passwords         + git push     + GitHub API

  ~6:05 AM   Scheduled Claude agent wakes up
     ↓       Pulls the cognitropy-library (private factory) and lab (public mirror)
     ↓       Runs cognitropy.py → gets today's (category, domain, technique, spark)
     ↓       Anti-clustering check against the last 5 days
     ↓       Copies the hybrid skeleton, then AUTHORS bespoke content
     ↓        (.claude/commands/, context/concepts/workflows/references, prompts)
     ↓       Scans for secrets leakage
     ↓       Commits to library; mirrors to lab
     ↓       Refreshes WORKSPACES.md + README badges + cognitropy-history.json
     ↓       Pushes lab; regenerates local dashboard HTML
     ↓       Cleans up the ephemeral deploy-path workspace
  Done.      One new workspace in the repo. Every day.

---

What's a "Workspace"?

Think of it as a ready-to-go AI assistant kit for a specific job. Each workspace is a folder you can point Claude Code (or any compatible AI CLI) at, and it instantly becomes an expert in that domain. It knows what questions to ask, what workflows to follow, and what commands are available.

Every workspace bundles bespoke domain commands (like /spectrum-sweep for an SDR workspace or /torque-split-design for a hybrid-powertrain workspace — never generic /analyze or /triage), substantive domain knowledge, methodology, reference tables, and reusable prompt templates. Clone one, launch your AI CLI inside it, and you're working.

Cognitropy workspaces follow a hybrid of two upstream Claude Code patterns: the directory shape from Claude-Workspace-Foundational-Plugin (flat context/, prompts/, outputs/ — no legacy ceremony), plus per-workspace .claude/commands/ for bespoke depth. The full spec is in WORKSPACE_SPEC.md.

workspace-name/
├── CLAUDE.md                      # Agent role, available commands, foundational instructions (~25-40 lines)
├── README.md                      # Human-facing showcase: what, why, quick start, refs (~50-100 lines)
├── .claude/
│   └── commands/                  # 4-8 bespoke domain commands (NO required /onboard)
│       └── <name>.md              # Each: brief purpose + Inputs / Steps / Output
├── context/
│   ├── concepts.md                # Domain knowledge — taxonomies, frameworks, failure modes (80-200 lines)
│   ├── workflows.md               # Methodology — step-by-step procedures, decision trees (50-150 lines)
│   └── references.md              # Lookup tables, cheat-sheets, links to upstream catalogues (30-100 lines)
├── prompts/                       # 2-4 bespoke reusable prompt templates
└── outputs/                       # Generated artifacts; .gitkeep at scaffold time

Workspaces are self-contained — no plugin install required. If you have the workspace-foundational plugin installed globally, additional cross-workspace commands (/workspace-foundational:context-sweep, /workspace-foundational:find-template) work alongside the workspace's own. Without it, the bespoke .claude/commands/ still work.

Older workspaces (project days 1–34, before 2026-05-10) follow a heavier legacy structure (/onboard interview command, context/{project,role,constraints}.md, context/for-agent/, planning/, work-log/, user-docs/, CREATION_REPORT.md) inherited from the retired Claude-Agent-Workspace-Model spec. They are kept as historical showcase artifacts — not retroactively migrated.

---

Quickstart

# Clone the lab
git clone https://github.com/DaxxSec/cognitropy-lab.git
cd cognitropy-lab

# Pick a workspace — any workspace
cd wireless-protocol-re   # or phishing-kit-analyzer, hybrid-system-energy-management, ...

# Open it as a Claude Code workspace
claude

Inside the workspace, the agent will read CLAUDE.md automatically and surface the bespoke commands defined in .claude/commands/. For wireless-protocol-re that means /spectrum-sweep, /decode-signal, /protocol-map, /compare-known, /report-findings, /capture-plan. For hybrid-system-energy-management that means /torque-split-design, /posterior-update, /ecms-tune, /wear-vs-fuel-pareto, etc. Every workspace publishes its own command set in its CLAUDE.md under ## Available Commands — no two workspaces share a command list, by design.

Each workspace is self-contained. The agent uses the repo as its memory — no cloud dependencies, no accounts to create, no API keys needed.

Browse all workspaces by category → see WORKSPACES.md.

---

Workspace Index

All current workspaces, grouped by category and updated daily by the build pipeline:

→ See WORKSPACES.md for the full catalog.

Each entry links to the workspace folder and includes its tagline. The index regenerates as part of every daily build, so it stays in sync with what's actually in the repo.

Engine Stats

Metric	Value
Cognitropy Domain Pool	363
Domain Categories	22
Technique Modifiers	30
Crossover Sparks	5
Crossover Probability	~30%
Total Unique Outcomes	18,863,790

The math: Standard days = 363 domains × 30 techniques = 10,890 combos. Crossover days = 125,686 cross-category domain pairs × 30 techniques × 5 sparks = 18,852,900 combos. Total: 18,863,790 unique possible assignments. At one workspace per day, that's 51,646 years before a repeat is even possible — and even then, the agent would build it differently. (Cross-category pairs = 363² − Σd² = 131,769 − 6,083 = 125,686 — because the engine enforces that crossover domains come from different categories.)

---

Run Your Own Lab

The cognitropy engine is generic. Fork the repo, swap in your own domain pool, point your own scheduled agent at it.

# Fork this repo, then edit the domain pool
vim cognitropy.py   # Replace DOMAINS list with your own interests

# Check what your custom pool generates
python3 cognitropy.py 2026-04-01
python3 cognitropy.py 2026-04-02
python3 cognitropy.py 2026-04-03

# Set up a scheduled Claude agent to build workspaces daily
# (see WORKSPACE_SPEC.md for structure conventions)

The workspace model works for literally anything. The domains are just the fun part.

---

About

The Cognitropy Lab is built by DaxxSec & Claude (Anthropic).

Inspired by Daniel Rosehill's Claude Code Projects Index and built on a hybrid of the (now-retired) Claude-Agent-Workspace-Model and the Claude-Workspace-Foundational-Plugin — see WORKSPACE_SPEC.md for the spec.

The daily build pipeline: Cognitropy assigns a domain → Claude agent builds the workspace → secrets scan → README stats update → push to GitHub → dashboard regeneration → local cleanup. Fully autonomous, every morning.

The term "cognitropic" and the underlying hash-based selection pattern were coined here. If you use it elsewhere, we'd love to hear about it.