Upstream status (2026-10-31): aawobdev/hermes-skills is no longer available. GitHub returns 404 and the repository cannot be cloned; the commands below are kept as published and will fail until the project reappears.

Blueprint Orchestration - Setup Guide¶

Source: aawobdev/hermes-skills Skills: 12 | Author: Alistair Hermes Version: Any (skills are standard SKILL.md format) Last Updated: June 22, 2026

Overview¶

A complete multi-agent project orchestration system for Hermes Agent. An expensive thinking model (Architect) interviews you, produces a detailed blueprint, then cheap local models execute role-by-role under your supervision.

Core philosophy: "The expensive model thinks. The cheap models do. You supervise."

Prerequisites¶

Hermes Agent installed and configured
Ollama (for local models) or OpenRouter API key (for cloud models)
Git (for cloning)
At least 24GB VRAM recommended for running multiple local models (see model-routing)

Quick Install¶

Method 1: Hermes Tap (Recommended)¶

hermes skills tap add aawobdev/hermes-skills

Skills are immediately available. Update anytime:

hermes skills tap update aawobdev/hermes-skills

Method 2: Manual Clone¶

git clone https://github.com/aawobdev/hermes-skills ~/hermes-skills

Add to ~/.hermes/config.yaml:

skills:
  external_dirs:
    - ~/hermes-skills/skills

Skills Included¶

Skill	Purpose	When to Use
`blueprint-orchestration`	Full workflow orchestrator	Start any new project
`role-architect`	System design, blueprint production	Phase 0-1: Interview + Planning
`role-designer`	UI/UX visual design	Phase 2: When UI is needed
`role-developer`	Code execution, file creation	Phase 3: Build
`role-tester`	Verification, drift detection	Phase 4: After each dev phase
`role-devops`	CI/CD, release, rollback	Phase 5: Deployment
`role-researcher`	Evidence gathering, option comparison	Any phase needing research
`role-security-auditor`	Vulnerability review	Phase 6: Before release
`role-end-user`	UX simulation	Phase 7: User-facing review
`role-orchestrator`	Task sequencing, one-shot delegation	Phase 3-7: Execution coordination
`prompting-standards`	LLM best practices	Reference for all role authors
`model-routing`	Model roster + assignments	Setup: pick models for your hardware

Starting a Project¶

In Hermes¶

/skill blueprint-orchestration

My project: I want to build a monitoring dashboard for my homelab.

Manual (Paste & Relay)¶

Open a session with a strong thinking model (e.g., qwen3.6-35b-a3b via LM Studio, or Claude Sonnet)
Paste the content of role-architect/SKILL.md
Provide your project in one sentence
Answer the 15 interview areas
Wait for the blueprint
Execute each phase using the assigned model

The Workflow¶

PHASE 0: Interview (Architect determines scope + roles)
    ↓
PHASE 1: Blueprint (Architect produces the full plan)
    ↓
PHASE 2: Design (Designer, if activated)
    ↓
PHASE 3: Build (Developer executes tasks)
    ↕
PHASE 4: Test (Tester after each dev phase)
    ↓
PHASE 5: Deploy (DevOps)
    ↓
PHASE 6: Security Audit
    ↓
PHASE 7: End-User review (if activated)
    ↓
PHASE 8: Closeout (Architect reviews, you sign off)

Any escalation → back to Architect
Human is the relay between all phases

Model Routing¶

Default Role-to-Model Assignments¶

Role	Primary Model	Runtime	tok/s
Architect	`qwen3.6-35b-a3b` (Q4_K_M)	LM Studio	~104
Designer	`qwen3.6-35b-a3b` + `gemma4:26b`	Mixed	varies
Developer	`qwen3-coder-30b` (Q4_K_M)	LM Studio	~135
Tester	`gemma4:26b` (Q4_K_M)	Ollama	~113
DevOps	`qwen3-coder-30b`	LM Studio	~135
Researcher	`qwen3.6-35b-a3b` or Claude Sonnet	Mixed	varies
Security	`qwen3.6-35b-a3b`	LM Studio	~104
End-User	`gemma4:26b`	Ollama	~113
Orchestrator	`deepseek/deepseek-v4-flash`	OpenRouter	cloud

Routing Tiers¶

TIER 0 - Orchestrator (OpenRouter)
  → Task decomposition, cheap reasoning

TIER 1 - Local Ollama (all routine dev, FREE)
  → Architect, Developer, Tester, End-User

TIER 2 - OpenRouter Paid
  → CC-class: complex logic, multi-file refactors

TIER 3 - OpenRouter Free (rate-limited)
  → Free cloud fallback

TIER 4 - Claude Code CLI (separate subscription)
  → CC-class via Claude: Sonnet/Haiku/Opus

Per-Project Setup¶

Each project needs a minimal HERMES.md:

# HERMES.md (in project root)

## Orchestration
Uses global Hermes skills. Run `/skill blueprint-orchestration` to start.
Blueprint is at: [path/to/blueprint.md]

## Project-specific model overrides
- Developer: qwen3-coder-30b (override from global)

## Project context for Hermes
- [One paragraph: what this project is, its stack, key constraints]

Customization¶

Role skills: Edit role-*/SKILL.md in place - changes take effect immediately
Model assignments: Edit model-routing/SKILL.md
Adding new roles: Add a new role-<name>/SKILL.md file
Prompt standards: Follow prompting-standards/SKILL.md Part A for blueprints, Part B for execution

Architect Interview Areas (15)¶

The Architect covers these during Phase 0:

Vision & outcome
Scope & boundaries
Platforms & targets
Technical environment
Stack & hosting selection
Execution context
Model inventory & routing
Risk & rollback
Quality & verification
Non-functional requirements
Design requirements
Security posture (shift-left)
User-facing surface
Deployment & operations
Patterns & preferences

Verification¶

After installation, verify skills are available:

# List installed skills
hermes skills list --marketplace | grep -E "blueprint|role-|model-routing|prompting"

# Load the main orchestrator
hermes chat
/skill blueprint-orchestration

Expected: The blueprint-orchestration skill loads and begins the Architect interview.

Troubleshooting¶

Skills not found¶

Verify external_dirs path in ~/.hermes/config.yaml is correct
Check directory permissions: ls -la ~/hermes-skills/skills/
Try re-installing: hermes skills tap update aawobdev/hermes-skills

Local models not loading¶

Check Ollama is running: ollama list
Verify model names match model-routing/SKILL.md
Pull missing models: ollama pull qwen3-coder:30b

VRAM exhaustion¶

Reduce concurrent loaded models
Use smaller quants (Q3_K_M instead of Q4_K_M)
Offload to OpenRouter for memory-heavy roles

← Skills Catalog | Blueprint Orchestration Marketplace Entry →

Powered by CorpusIQ

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects