Hermes Profile Packs

August 24, 2026 · View on GitHub

Curated profile packs for Hermes Agent.

This repository is the public upstream home for reusable Hermes profile distributions maintained as coherent teams rather than isolated prompts.

Packs

Pack	Namespace	Purpose
Hermes Agency	`agency-*`	Professional specialists for engineering, product, design, QA, operations, marketing, content, and coordinated project work.
Hermes Council	`council-*`	Personal specialists for growth, health, fitness, parenting, relationships, communication, community, faith, finances, home stewardship, personal admin, career, learning, recreation, travel, nutrition, sleep, and preparedness.
Hermes Academy	`academy-*`	Professors and instructors for academic, technical, vocational, and professional learning.

The packs deliberately separate durable context domains. Agency executes professional work, Council supports the person, and Academy teaches. A work profile should not need private life context, a personal profile should not inherit repository state merely because both run on Hermes, and a teaching profile should not silently become a production executor.

Hermes Council currently contains 21 focused profiles and 84 purpose-built personal-life skills. Hermes Academy v0.2 contains 30 faculty profiles and 120 purpose-built teaching skills. Hermes Agency contains 109 professional specialists.

Academy Continuing Education

Hermes Academy includes a Continuing Education flow for profile-to-profile competency transfer. A learner can receive a natural-language request such as Go learn API security, route through the real Academy Dean when the user did not name a teacher, establish native Hermes goal state, study with the selected faculty member through canonical Bot messaging, prove transfer, and persist a reusable capability through native /learn into the learner's normal skill store. It adds no custom training runtime or Fleet dependency.

The implementation passed Phase 15 independent production-profile verification and the Phase 16 whole-change/release review on the maintained Dadmin88/hermes-agent-downstream integration. It is production-validated on that maintained downstream path. Stock NousResearch Hermes compatibility remains pending until the normal Hermes release provides equivalent generic support for the native seams required by the validated flow, so do not read this as an unqualified stock-Hermes readiness claim.

Phase 16 also hardened Dean routing: category fallbacks are manifest-owned, cross-category CE objectives fail closed instead of falling into an unrelated broad chair, unavailable specialists may use only a safe installed fallback from their own category, and the installed Dean's preloaded routing contract no longer depends on a pack-root Python helper that is absent after normal profile installation.

See docs/CONTINUING_EDUCATION.md, the normative docs/CONTINUING_EDUCATION_ARCHITECTURE.md, and the Phase 16 release review.

Start small with Team Recipes

You rarely need all 160 profiles. Team Recipes are validated compositions of existing profiles for common outcomes, with minimal, recommended, and expanded tiers.

The main installer understands them directly:

python install.py --list-recipes
python install.py --recommend "build and ship a web app"
python install.py --recipe software-delivery --tier minimal --dry-run
python install.py --recipe software-delivery --tier recommended --yes

Recommendation is confidence-aware. A recipe is promoted only when one formation is a strong, unambiguous fit; otherwise the installer falls back to individual profile matches instead of forcing a template.

Recipes are portable selection guidance only. They do not create memory, cron jobs, Bot groups, Kanban state, credentials, or other runtime state. Installation still uses the normal Profile Packs/Hermes distribution path.

See docs/TEAM_RECIPES.md and docs/FIRST_TEAM.md.

Repository layout

hermes-profile-packs/
├── hermes-agency/
├── hermes-council/
├── hermes-academy/
├── docs/
│   ├── CONTINUING_EDUCATION.md
│   ├── CONTINUING_EDUCATION_ARCHITECTURE.md
│   └── CONTINUING_EDUCATION_RELEASE_REVIEW.md
├── examples/
├── tests/
├── packs.json
├── recipes.json
├── recipe_catalog.py       # Shared recipe loading/scoring/confidence engine
├── install.py              # Canonical human + agent profile/recipe selector
├── recipes.py              # Focused recipe-only convenience client
└── validate.py

Runtime state is never distributed. Authentication material, .env files, logs, caches, databases, local paths, session history, and machine-specific configuration do not belong in this repository.

Install

For most people, launch the selector:

python install.py

The wizard can recommend a validated Team Recipe when a goal maps cleanly to one, browse recipes directly, browse packs/categories, search exact profiles, preview the exact install plan, and still offer an explicit install-everything path.

For agents and scripts:

python install.py --agent-help --json
python install.py --catalog --json
python install.py --list-recipes --json
python install.py --recommend "build a web app" --json
python install.py --recipe software-delivery --tier minimal --dry-run --json
python install.py --recipe software-delivery --tier recommended --yes --json
python install.py --profiles agency-backend-engineer agency-frontend-engineer --dry-run --json
python install.py --all --yes --json

--catalog, --list-recipes, and --recommend are read-only. Existing --recommend --json consumers keep the individual-profile recommendations array; recipe results are additive through recipe_match and recipe_recommendations.

Legacy pack-first commands remain supported.

See docs/INSTALLER.md for the full contract.

Focused recipe client

recipes.py remains supported when a recipe-only surface is convenient:

python recipes.py --list
python recipes.py --recommend "learn cybersecurity"
python recipes.py cybersecurity-learning --tier recommended --dry-run

Both entry points use the same shared recipe engine.

Operate the team

Use docs/OPERATING_PLAYBOOK.md, docs/INSTALLATION_VERIFICATION.md, and docs/TROUBLESHOOTING.md.

Validate

python validate.py
python -m unittest discover -s tests -p 'test_*.py'

This validates pack manifests, namespaces, distribution metadata, skills, portability, secret/path hygiene, team recipe references/tier invariants, and root selector behavior.

Design principles

Coherent teams, not prompt dumps. Profiles have explicit ownership and clean handoffs.
Namespace isolation. agency-* is professional, council-* is personal, and academy-* is educational.
Portable distributions only. No runtime state, secrets, personal filesystem paths, or local caches.
Specialists remain specialists. Profiles hand off rather than silently absorbing unrelated domains.
Smallest useful team. Add profiles only for distinct expertise, independent review, or useful parallelism.
Conservative recommendation. A recipe is promoted only when the deterministic match is strong and unambiguous.
Human agency stays central. Council supports decisions and capability; it does not run a person's life.
Safety beats role-play. Profiles do not manufacture authority they do not possess.
Academy teaches for transfer. Teaching should make the learner progressively more capable.

See PACK_SPEC.md, CONTRIBUTING.md, and SECURITY.md.

License

AGPL-3.0-only. See LICENSE.