Great Minds

14 legendary personas (Jobs, Musk, Buffett, Ive, Rubin, Huang, Winfrey, Rhimes, Blakely, Hamilton, Angelou, Sorkin, Aurelius, Jackson) plus 8 functional-role implementers — that debate, review, plan, and build. Available in three install formats to match how you work.

📖 Read the User Manual for the complete reference — install paths, all 22 agents (14 named + 8 functional), all 17 skills, lite vs full agency workflows, project structure, memory store, three end-to-end walkthroughs, patterns, and troubleshooting.

Newcomer? Start here

/constellation-start

The discoverable entry point for the Caseproof persona constellation. Asks 2-3 questions about your project shape and routes to the right plugin's project-init skill — or channels Phil Jackson (constellation-aware as of v1.3) for ambiguous or mixed projects. Always ensures the bible at .great-authors/project.md exists, since that's the shared spine across all five plugins.

If you already know which plugin you want, you can skip the entry point and run that plugin's project-init directly. /constellation-start exists for the case where you don't.

What's new in v1.4

Polish from the cross-plugin quality review. No new user-facing features; everything is hygiene and parity with the rest of the constellation.

Added

tests/smoke.sh — first smoke test for great-minds. Validates SKILL.md frontmatter, persona frontmatter (name / description / model / color), DXT persona-subset membership against agents/, version coherence across package.json / plugin.json / DXT manifests / DXT server self-report, DXT tool-handler alignment, and v1.3+ /constellation-start skill presence.
LICENSE — MIT, matching the rest of the constellation. The plugin had been shipping without one.
Frontmatter added to three SKILL.md files that predated the convention: agency-daemon, scope-check, agency-memory.
color: field added to aaron-sorkin-screenwriter.md (the only persona missing it).

Changed

distribution/plugin/skills/constellation-start/ — the lite distribution now includes the constellation entry point. Newcomers using the lite install can run /constellation-start without falling back to the full bundle. Phil Jackson's persona file in lite is also resynced with the constellation-aware expansion from v1.3.

Why

A cross-plugin review pass surfaced these as accumulated drift. None were user-blocking, but together they were a real consistency gap — every other plugin in the constellation had smoke tests, and great-minds was the most likely to drift without them given its breadth (22 personas, 20 skills, infrastructure for a swarm runtime).

What's new in v1.3

Constellation-level entry point. The plugin now has a discoverable starting point for newcomers to the entire Caseproof persona constellation (great-minds + great-authors + great-filmmakers + great-publishers + great-marketers — five plugins, each owning one craft).

/constellation-start (new skill) — asks project-shape questions, routes to the right plugin's project-init or channels Phil Jackson for mixed projects. Discoverable, lightweight, ~150 lines.
phil-jackson-orchestrator.md (expanded) — Phil's existing role (Zen-master swarm orchestrator within great-minds) is preserved unchanged; a new "Your Role at the Constellation Level" section adds the orchestration knowledge for cross-plugin work. Same triangle-offense logic; one level up. Eleven rings, not one.
DXT bundle — adds the constellation_start tool. Available in Claude Desktop alongside the existing tools.
Version coherence — DXT manifests caught up to plugin.json baseline (was 1.0.0; now 1.3.0 to match). Cleaner state for future smoke-test infrastructure.

Doesn't break anything. /agency-start and the entire agency-* swarm-pattern family stay intact — they're the right answer for software-product swarm projects, and /constellation-start dispatches to them when the user picks that branch.

What's new in v1.2

The three-layer architecture is now explicit and shipping. Grounded in 2026 research (Wharton, USC) on persona prompting — named expert personas excel at judgment but reduce factual accuracy on knowledge-heavy tasks; functional roles avoid that trap.

Layer 1 — Named directors (Sonnet): Jobs, Musk, Hamilton, Aurelius, the Board. Vision, judgment, mediation.
Layer 2 — Named specialists (Sonnet): Ive, Angelou, Rubin, Sorkin, Blakely. Domain craft with character.
Layer 3a — Functional code-writers (Sonnet, NEW): backend-engineer, frontend-developer, database-architect, devops-engineer, test-engineer. Code that integrates with a real codebase needs Sonnet's reasoning even without a named voice.
Layer 3b — Functional reviewers / doc-writers (Haiku, NEW): code-reviewer, security-auditor, documentation-writer. Pattern-matching and recall work where Haiku's speed and cost win.

The directors at Layer 1 now know which kind of work to dispatch to which layer — codified in their persona files and in AGENTS.md. Steve and Elon dispatch to Layer 2 for craft and Layer 3 for code; Margaret dispatches to Layer 3 for QA work; Aurelius coordinates across layers. Three-layer model documented in MANUAL.md Section 4 and docs/OPERATIONS.md.

Recommended companion: Superpowers

Great Minds and the Superpowers plugin operate at different altitudes and compose well:

GSD (built into Great Minds) — the agency-level pipeline: PRD → debate → plan → execute → verify → ship. Coordinates which kind of work happens when across multiple agents over a multi-day project.
Superpowers — the engineer-level discipline within each task: TDD, systematic debugging, plan-then-execute, worktree isolation, pre-merge verification, code review patterns.

Use them together: GSD orchestrates the agency; Superpowers gives the Layer 3 code-writers and reviewers the engineering discipline they need within each dispatch. The directors at Layer 1 don't follow Superpowers TDD discipline (their work isn't TDD-shaped), but the backend-engineer, frontend-developer, test-engineer, code-reviewer, and security-auditor agents do.

Specific composition:

GSD pipeline phase	Apply Superpowers skill
`/agency-plan`	`writing-plans`, `brainstorming`
`/agency-execute`	`subagent-driven-development`, `using-git-worktrees`, `dispatching-parallel-agents`
`/agency-verify`	`verification-before-completion`, `systematic-debugging`
Layer 3 code-writer dispatch	`test-driven-development`, `requesting-code-review`
Pre-merge	`receiving-code-review`, `finishing-a-development-branch`

Install Superpowers if you want the engineering-discipline layer beneath the agency-pipeline layer. Both plugins keep their own identities and update cadences — they're composed, not merged.

Pick your format

Format	Install target	Best for	Where
`great-minds` (full)	Claude Code	Power users running the autonomous PRD → ship pipeline (cron, daemon, worktrees)	this repo root
`great-minds-lite`	Claude Cowork + Claude Code	Teams that want personas + co-work skills (debate, board review, plan) without the swarm	`distribution/plugin/`
DXT bundle	Claude Desktop app	Non-technical teammates, one-click install	`distribution/dxt/`

All three share the same 14 personas. Edit once in agents/ at repo root — a pre-commit hook syncs to the other targets.

Install

Full agency (Claude Code):

/plugin marketplace add sethshoultes/great-minds-plugin
/plugin install great-minds@sethshoultes

Lite (Cowork or Code):

/plugin marketplace add sethshoultes/great-minds-plugin
/plugin install great-minds-lite@sethshoultes

DXT (Desktop app):

cd distribution/dxt && npm install && npx @anthropic-ai/dxt pack
# Share the generated great-minds.dxt — teammates double-click to install.

See distribution/README.md for full details on the lite + DXT options.

What You Get

(The sections below document the full great-minds agency. For the slim lite plugin or DXT tool list, see distribution/README.md.)

What You Get

14 Agent Personas + 2 Internal Consolidation Functions

Agent	Role
`phil-jackson-orchestrator`	Orchestrator — system coordination, dispatch, resource optimization
`steve-jobs-visionary`	Design & Brand — simplicity, taste, the human experience
`elon-musk-persona`	Product & Growth — first principles, feasibility, scale
`jensen-huang-board`	Board Member — tech strategy, data moats, platform economics
`oprah-winfrey-board`	Board Member — audience connection, storytelling, accessibility
`warren-buffett-board`	Board Member — business model, unit economics, moat durability
`shonda-rhimes-board`	Board Member — narrative & engagement, retention loops, cliffhangers
`margaret-hamilton-qa`	QA Director — zero-defect methodology, continuous testing
`rick-rubin-creative`	Creative Director — strip to essence, authenticity
`jony-ive-designer`	Visual Design — spacing, hierarchy, craft, inevitability
`maya-angelou-writer`	Copywriting — warmth, rhythm, dignity, emotional resonance
`aaron-sorkin-screenwriter`	Screenwriter — video scripts, demos, tutorials, launch videos
`sara-blakely-growth`	Growth Strategy — scrappy, customer-first, grassroots
`marcus-aurelius-mod`	Moderator — Stoic orchestration, conflict mediation

The daemon pipeline also uses 2 internal consolidation functions (not standalone agents):

philJacksonConsolidation — Merges debate decisions into a single blueprint after Round 2
boardConsolidation — Consolidates all 4 board member reviews into a unified verdict

17 Skills (Slash Commands)

Command	Description
`/agency-start <name>`	Initialize a new agency project with full system files, dirs, memory, and crons
`/agency-launch`	Launch the pipeline: debate → plan → execute → verify → ship
`/agency-status`	Comprehensive health check — pipeline, crons, memory, server, GitHub, commits
`/agency-debate <topic>`	Structured 2-round Steve vs. Elon debate with Rick Rubin essence check
`/agency-plan`	GSD-style structured task planning with XML plans
`/agency-execute`	Wave-based parallel execution with fresh context per task
`/agency-verify`	Automated UAT verification + QA pipeline (build, a11y, security, live site)
`/agency-board-review`	Full board review — Jensen, Oprah, Warren, Shonda review in parallel, then consolidate
`/agency-ship`	Ship phase — merge, deploy, retrospective, scoreboard update
`/agency-crons`	Install decoupled cron system (bash + haiku, no bottleneck)
`/agency-memory`	Memory operations — store, recall, consolidate learnings
`/agency-publish`	Publish deliverables to external platforms
`/agency-video`	Generate video scripts and storyboards
`/agency-daemon`	Long-running Agent SDK daemon -- continuous orchestration, replaces cron pipeline
`/agency-anatomy`	File anatomy -- token estimates per file for context budgeting
`/agency-tokens`	Token ledger -- cost tracking per agent across pipeline runs
`/scope-check`	Detect scope creep against original plan

Hooks

SubagentStop — Reminds directors to review output and not go idle
Context Guard — Warns when context is getting large, suggests fresh agents

Templates

Project system files (SOUL.md, AGENTS.md, TASKS.md, STATUS.md, etc.)
.planning/ directory (GSD-style: PROJECT.md, REQUIREMENTS.md, ROADMAP.md, STATE.md)
PRD template
Agent hiring template

The Pipeline

PRD → Debate (2 rounds) → /agency-plan → /agency-execute → /agency-verify → Ship

Each phase uses the right tool:

Debate: Agent tool with worktree isolation (Steve + Elon in parallel)
Plan: Structured XML task plans verified against PRD requirements
Execute: Wave-based parallel agents, each with fresh context + isolated worktree
Verify: Margaret QA — build, lint, tests, requirement coverage
Ship: Merge, update scoreboard, memory write

Architecture

You (Phil Jackson — Orchestrator)
  ├── Board of Directors (parallel via /agency-board-review)
  │    ├── Jensen Huang — Tech Strategy (cron reviews, GitHub issues)
  │    ├── Oprah Winfrey — Audience & Accessibility
  │    ├── Warren Buffett — Business & Economics
  │    └── Shonda Rhimes — Narrative & Engagement
  ├── Agent tool (worktree) → Steve Jobs — Creative Director
  │    └── Agent tool (haiku) → Jony Ive, Maya Angelou, Rick Rubin
  ├── Agent tool (worktree) → Elon Musk — Product Director
  │    └── Agent tool (haiku) → Sara Blakely + engineering sub-agents
  └── Agent tool (worktree) → Margaret Hamilton — QA Director

Key insight: Agent tool with worktree isolation is the reliable dispatch method. Each agent gets an isolated copy of the repo, creates a branch, builds, commits, pushes. No conflicts, no context rot.

What does NOT work

tmux send-keys — Claude Code's input buffer rejects pasted prompts. Workers sit idle.
Cron dispatch via tmux — can't reliably send commands to other terminals
In-conversation crons — bottleneck the main agent

Daemon (Primary Orchestration)

The daemon (/agency-daemon) is an Agent SDK-based long-running process that replaces the cron pipeline. It handles dispatch, health checks, dream consolidation, and memory maintenance in a single persistent process.

Daemon Sync

⚠️ TEMPORARY TOOL — Scheduled for deprecation once daemon is extracted to shared package.

This tool syncs daemon code from great-minds-plugin to the great-minds repository. It is a surgical band-aid until the daemon becomes a standalone npm package.

First Time Setup

Edit hardcoded paths in sync.sh (lines 11-12) for your environment:

PLUGIN_ROOT="/Users/sethshoultes/Local Sites/great-minds-plugin"
GREAT_MINDS_ROOT="/home/agent/great-minds"

Examples:

macOS: /Users/yourname/path/to/repo
Linux: /home/yourname/path/to/repo
Docker: /app/repo

Usage

Run after editing daemon code in the plugin repo:

npm run sync

What it does:

Validates destination repo has no uncommitted changes
Copies 11 TypeScript files from daemon/src/
Copies daemon/package.json and daemon/README.md
Copies BANNED-PATTERNS.md and DO-NOT-REPEAT.md
Updates CLAUDE.md with sync protocol and anti-hallucination rules
Runs npm install in destination
Commits and pushes changes

Execution time: <5 seconds (excluding git push network time)

Binary outcome: Script exits 0 on success, 1 on failure. No partial syncs.

Deprecation Timeline

3 months: Extract daemon to shared npm package
6 months: Delete sync.sh if extraction complete, otherwise escalate as architectural debt

Daemon Resilience

The daemon includes production-grade resilience features:

Telegram Notifications — Real-time alerts for pipeline starts, completions, failures, and hung agents. Requires TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID environment variables.
Crash Recovery — Failed pipeline phases retry up to 2 times with exponential backoff. If all retries fail, the PRD is archived to prds/failed/ so it does not block the queue.
Hung Agent Detection — Individual agents timeout after 20 minutes (AGENT_TIMEOUT_MS). The entire pipeline timeout is 60 minutes (PIPELINE_TIMEOUT_MS). Hung agents are killed and the phase is retried or skipped.

Telegram Setup

Message @BotFather on Telegram and create a new bot (/newbot)
Copy the bot token
Send a message to your bot, then fetch your chat ID via https://api.telegram.org/bot<TOKEN>/getUpdates

Set environment variables:

export TELEGRAM_BOT_TOKEN="your-bot-token"
export TELEGRAM_CHAT_ID="your-chat-id"

Developer Intelligence

Inspired by OpenWolf, the daemon includes developer intelligence features:

File Anatomy (/agency-anatomy) — Token estimates per file, helping agents budget context windows and avoid loading oversized files.
Token Ledger (/agency-tokens) — Tracks token usage and cost per agent across pipeline runs. Shows which agents are expensive and where to optimize.
Bug Memory — 8 known bugs stored in a searchable buglog (daemon/buglog.json). Agents query this before debugging to avoid re-investigating known issues.
Do-Not-Repeat List — A list of past mistakes and anti-patterns (daemon/do-not-repeat.json) injected into every agent session to prevent regression.

Legacy Cron System (Fallback)

Crons are still available via /agency-crons for environments where the daemon cannot run:

Cron	Model	Cost
Heartbeat (5 min)	Bash	Free
QA checks (29 min)	Bash	Free
Git monitor (15 min)	Bash	Free
Dispatch (30 min)	Haiku	Cheap
Dream consolidation (60 min)	Haiku	Cheap

Prerequisites

Claude Code CLI
git

That's it. No tmux, no claude-swarm, no external dependencies.

Quick Start

Option 1: Claude Code Plugin (Interactive)

Use the plugin directly inside Claude Code for interactive, command-driven builds.

# 1. Install the plugin
npx plugins add sethshoultes/great-minds-plugin

# 2. Start a new project (creates SOUL.md, AGENTS.md, TASKS.md, STATUS.md, prds/)
/agency-start my-product

# 3. Write a PRD
#    Create prds/my-product.md describing what you want built.
#    Use prds/TEMPLATE.md as a starting point.

# 4. Launch the pipeline (one-shot, runs all phases in sequence)
/agency-launch

# 5. Or run individual phases
/agency-debate "Should we use React or Astro?"
/agency-plan
/agency-execute
/agency-verify
/agency-board-review
/agency-ship

# 6. Check status anytime
/agency-status

Option 2: Daemon (Autonomous)

Run the daemon for continuous, autonomous builds. Drop PRDs in and walk away.

# 1. Clone the plugin and your target repo
git clone https://github.com/sethshoultes/great-minds-plugin.git
git clone https://github.com/your-org/your-project.git

# 2. Install daemon dependencies
cd great-minds-plugin/daemon
npm install

# 3. Start the daemon pointed at your repo
PIPELINE_REPO=/path/to/your-project npx tsx src/daemon.ts

# 4. Drop a PRD in your repo's prds/ directory
cp prds/TEMPLATE.md /path/to/your-project/prds/my-feature.md
# Edit the PRD with your requirements

# 5. The daemon automatically:
#    - Detects the new PRD
#    - Runs debate (Steve vs Elon, 2 rounds + Rick Rubin + Phil Jackson)
#    - Plans (task breakdown + Sara Blakely gut check)
#    - Builds (code, tests, commits)
#    - QA (2 passes by Margaret Hamilton — including live testing)
#    - Creative review (Jony Ive, Maya Angelou, Aaron Sorkin)
#    - Board review (Jensen, Oprah, Buffett, Shonda)
#    - Ships (commit, push, merge to main, retrospective)
#    - Archives the PRD to prds/completed/

Option 3: Docker

cd great-minds-plugin/daemon
docker build -t greatminds-daemon .
docker run -d \
  -e PIPELINE_REPO=/repo \
  -v /path/to/your-project:/repo \
  greatminds-daemon

Option 4: Server (systemd)

For always-on autonomous builds on a remote server:

# Create the service
sudo cat > /etc/systemd/system/greatminds-daemon.service << 'EOF'
[Unit]
Description=Great Minds Pipeline Daemon
After=network.target

[Service]
Type=simple
User=your-user
WorkingDirectory=/path/to/great-minds-plugin/daemon
Environment=PIPELINE_REPO=/path/to/your-project
Environment=HOME=/home/your-user
ExecStart=/usr/bin/npx tsx src/daemon.ts
Restart=always
RestartSec=10
StandardOutput=append:/tmp/claude-shared/daemon.log
StandardError=append:/tmp/claude-shared/daemon.log

[Install]
WantedBy=multi-user.target
EOF

sudo systemctl daemon-reload
sudo systemctl enable greatminds-daemon
sudo systemctl start greatminds-daemon

Writing a Good PRD

Your PRD is the input that drives everything. A good PRD includes:

Problem — What problem does this solve?
Solution — What are you building?
Requirements — Specific, testable requirements
Technical constraints — Framework, language, deployment target
Success criteria — How do we know it's done?

See prds/TEMPLATE.md for the full template.

Environment Variables

Variable	Default	Description
`PIPELINE_REPO`	`../../`	Target repo the daemon builds in
`AGENT_TIMEOUT_MS`	`1200000` (20 min)	Max time per agent call
`PIPELINE_TIMEOUT_MS`	`3600000` (60 min)	Max time per pipeline run
`TELEGRAM_BOT_TOKEN`	(none)	Telegram notifications
`TELEGRAM_CHAT_ID`	(none)	Telegram chat ID

Related Projects

great-minds — The agency repo
shipyard-ai — Autonomous site builder (spun out)
localgenius — First product built by the agency

Contributor setup

One-time, after cloning:

git config core.hooksPath .githooks

This enables a pre-commit hook that auto-syncs agents/*.md edits into distribution/plugin/ and distribution/dxt/ so the three formats never drift. Manual sync: ./distribution/sync-distribution.sh.

License

MIT

Name		Name	Last commit message	Last commit date
Latest commit History 178 Commits
.claude-plugin		.claude-plugin
.githooks		.githooks
agents		agents
crons		crons
daemon		daemon
distribution		distribution
docs		docs
hooks		hooks
memory-store		memory-store
skills		skills
templates		templates
tests		tests
.daemon-queue.json		.daemon-queue.json
.github-intake-state.json		.github-intake-state.json
.gitignore		.gitignore
AGENT-TRIGGERS.md		AGENT-TRIGGERS.md
BANNED-PATTERNS.md		BANNED-PATTERNS.md
DEPRECATED.md		DEPRECATED.md
DO-NOT-REPEAT.md		DO-NOT-REPEAT.md
LICENSE		LICENSE
MANUAL.md		MANUAL.md
README.md		README.md
package.json		package.json
sync.sh		sync.sh

Folders and files

Latest commit

History

Repository files navigation

Great Minds

Newcomer? Start here

What's new in v1.4

Added

Changed

Why

What's new in v1.3

What's new in v1.2

Recommended companion: Superpowers

Contents

Pick your format

Install

What You Get

What You Get

14 Agent Personas + 2 Internal Consolidation Functions

17 Skills (Slash Commands)

Hooks

Templates

The Pipeline

Architecture

What does NOT work

Daemon (Primary Orchestration)

Daemon Sync

First Time Setup

Usage

Deprecation Timeline

Daemon Resilience

Telegram Setup

Developer Intelligence

Legacy Cron System (Fallback)

Prerequisites

Quick Start

Option 1: Claude Code Plugin (Interactive)

Option 2: Daemon (Autonomous)

Option 3: Docker

Option 4: Server (systemd)

Writing a Good PRD

Environment Variables

Related Projects

Contributor setup

License

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases 4

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages