Discovery Orchestrator

Multi-expert discovery & needfinding skill for Claude Code & Hermes Agent. Transforms fuzzy ideas into decision-ready artifacts using parallel expert agents.

What It Does

Discovery Orchestrator helps you frame problems before jumping to solutions. It runs structured discovery conversations enhanced by 9 parallel expert agents, each analyzing your situation from a specialized perspective:

Expert	Perspective
Facilitator	Pacing, flow, cognitive load
Needfinding	Pain signals, workarounds, latent needs
JTBD Analyst	Jobs To Be Done — functional, emotional, social
Requirements Eng	Constraints, assumptions, scope boundaries
Product Researcher	Opportunity scoring, HMW questions, hypotheses
Psychologist	Cognitive biases, emotional signals
MI Coach	Motivational interviewing, change readiness
Systems Advisor	Feasibility, dependencies, risk
Decision Analyst	Trade-offs, prioritization, reversibility

Zero dependencies — works with Claude Code CLI or Hermes Agent. No Python, no MCP servers, no plugins.

Installation

Claude Code (default)

git clone https://github.com/doz34/discovery-orchestrator.git
cd discovery-orchestrator
chmod +x install.sh
./install.sh

Or manually:

git clone https://github.com/doz34/discovery-orchestrator.git
cp -r discovery-orchestrator ~/.claude/skills/discovery-orchestrator

Hermes Agent

git clone https://github.com/doz34/discovery-orchestrator.git
cd discovery-orchestrator/hermes
chmod +x install.sh
./install.sh

Or manually:

git clone https://github.com/doz34/discovery-orchestrator.git
# Copy shared content + Hermes SKILL.md
mkdir -p ~/.hermes/skills/discovery-orchestrator
cp discovery-orchestrator/hermes/SKILL.md ~/.hermes/skills/discovery-orchestrator/
for dir in experts phases adapters templates scoring examples; do
  cp -r discovery-orchestrator/$dir ~/.hermes/skills/discovery-orchestrator/$dir
done

Usage

Claude Code

Type:

/discovery-orchestrator I want to build a RAG system for legal documents

Hermes Agent

Invoke:

discovery-orchestrator I want to build a RAG system for legal documents

Or just start describing your situation — the skill auto-detects domain and complexity on both platforms.

Examples

Claude Code:

/discovery-orchestrator Our team spends 3 hours/week on manual report generation

Hermes Agent:

discovery-orchestrator Our team spends 3 hours/week on manual report generation

How It Works

Input → Phase 0: Orientation (domain + complexity detection)
      → Phase 1-2: Conversation (concrete case + friction mapping)
      → Phase 3+: Parallel expert agents (spawned via Agent/delegate tool)
      → Synthesis + Quality scoring (5-dimension rubric)
      → Output artifact (saved to .discovery/)

Architecture

The skill uses lazy loading — only the phases and experts needed for your situation are loaded. This keeps token usage at 3-6K vs the 15-20K a monolithic approach would consume.

Modes

Mode	When	Route	Experts
Rapid	Already clear, want quick validation	P1→P3→P7→P9	2
Structured	Real problem, needs framing	P1→P2→...→P9	5
Deep	Ambiguous, strategic, emotional	Full + all phases	9

Domains

• Software — code, repos, APIs, features, architecture
• Product/UX — users, journeys, retention, conversion
• AI Projects — models, agents, prompts, RAG systems
• Consulting — stakeholders, diagnostics, proposals
• Internal Tools — workflows, automation, ops
• Personal — career decisions, life choices

Output Templates

Template	Use Case
Concise Framing	Rapid validation
Discovery Brief	Default structured output
Agent Handoff	Ready for downstream AI agent
Spec Pre-Brief	Bridge to formal specification
Stakeholder Matrix	Multi-stakeholder alignment

Quality Scoring

Every discovery is scored on 5 dimensions before delivery:

Dimension	Weight	Measures
Specificity	25%	Vague wishes → measurable criteria
Actionability	25%	No direction → decision-ready
Traceability	20%	Unsourced → every claim traced to user
Completeness	15%	Blind spots → all angles explored
Coherence	15%	Contradictions → aligned narrative

Passing threshold: 3.5/5.0. Below threshold triggers an additional clarification loop.

Extending

Add a Domain Adapter

Create adapters/your-domain.md following the existing format. The orchestrator auto-detects new adapters.

Add an Expert

Create experts/your-expert.md with the standard format (Role, Activation triggers, Analysis framework, Probes, Output format, Red flags).

Add a Template

Create templates/your-template.md. Reference it in Phase 9 handoff.

Project Structure

discovery-orchestrator/
├── SKILL.md              # Claude Code orchestrator (~140 lines, always loaded)
├── hermes/
│   ├── SKILL.md          # Hermes Agent orchestrator (native delegate tool)
│   └── install.sh        # Hermes installer
├── experts/              # 9 expert agents (loaded on demand, shared)
├── phases/               # 10 phases (loaded per route, shared)
├── adapters/             # 6 domain adapters (shared)
├── templates/            # 5 output templates (shared)
├── scoring/              # Quality rubric (shared)
├── examples/             # Example sessions
├── install.sh            # Claude Code installer
└── LICENSE               # MIT

Comparison: v1 vs v2

Metric	v1	v2
Files	1 monolith	41+ modular
Expert depth	~15 lines/role	120-250 lines/agent
Real agents	0	3-9 parallel
Tokens/load	15-20K	3-6K (lazy)
Quality gate	None	5D rubric
Persistence	None	Markdown files
Extensibility	None	Add files = done

License

MIT — see LICENSE.