The Breeze
Private · Empire Edition
Hash check · Session in localStorage

Empire Skills Library · Edition 1 · August 2026

Four hundred
and twenty-four
small machines.

A skill is a procedure, encoded as a markdown file, that one operator or agent can load, follow, and verify. The B3RT empire runs on — of them. This page is a guided tour — what they are, who owns them, and how to invoke the dozen that earn their place in the daily rotation.

live total skills 8 categories 20+ legion agents 8 domain families

01

Why a skills library at all

The same answer every other operator eventually finds: because documentation that lives in chat dies when the chat scrolls, and because a skill is the smallest artifact that survives context loss.

A skill is not a notebook entry and not a prompt template. A skill is a contract. It has a name, a one-line description, a trigger condition, a procedure, references, and a verification. When the same procedure gets invoked three times across three sessions, the third invocation should not start from zero — it should load the skill, run its procedure, and report what changed since the last run.

The empire's 424 skills live under ~/AppData/Local/hermes/skills/ as plain SKILL.md files. Each file is human-readable, agent-loadable, and self-contained. There is no index database to keep in sync — the file system is the index, and the Empire Hub exposes it at /api/skills with a 60-second disk cache. If a skill exists on disk, it is in the library. If it has been deleted, it is gone. The library never lies about itself.

The 1500-word version. The empire's edge is not its models, its stack, or its domains. It is the library. When an agent at 02:00 in the morning receives a task it has never seen, the library is what lets it pick up a procedure that a human ran at 14:00 in the afternoon three weeks ago, follow it, and produce an artifact the human would have produced — only faster, and without the mental load of remembering how. The library is the only piece of the empire that gets more valuable the longer it sits there.

Skills get written when a procedure recurs. They get pruned when a procedure stops recurring. There is no top-down skill taxonomy meeting — there is only the operator noticing that they did the same five things three times in a row, and writing down the five things so the fourth time takes a tenth of the effort. The library is therefore a by-product of operation, not a deliverable in itself. That is the only way a one-person library scales past fifty entries without rotting.

02

Categories

Eight domain families. The boundaries are rough by design — a skill that lives in "Empire" is one that ships value inside the empire, and a skill that lives in "Productivity" is one that operates on the operator's own work loop.

Legion

—

The named agent personas. Scholar-Omega, Builder-Sigma, Cypher, Fixer, Guardian, Bard, Ted Lasso, Scribe, Warhol, Storyteller, Sparks, Maverick, and the rest. Each persona ships its own skills, its own deliverables, and its own accountability surface.

Empire

—

Operating skills for the empire itself. Phalanx site probes, pulse aggregation, hub launch, sitemap generation, drift detection, empire grade computation, cron orchestration, supervisor restart, knowledge graph build.

Productivity

—

Operator-level work. Kanban triage, morning standup, handoff compactor, error handler, document-to-action-items, imperator brief, evening standup, spoken empire briefing.

Security

—

Defensive skills. Dependency audit, secret scan, security headers rollout, Tailscale remote access, OAuth setup, validator-iotas. Treat these as load-bearing — every FastAPI service ships SecurityHeadersMiddleware and these skills are why.

AI / ML

—

Model ops. Llama.cpp local inference, HuggingFace hub, Weights & Biases, audio/video model skills, segment anything, TTS, and the small library of skills that make a local 35B model behave like a service.

Business

—

BE Repaired ops, Sparkling Solutions booking flow, lead enrichment, quote builder, invoice chaser, the BE Repaired supervisor, and the AI Research Lab LSB surface.

Research

—

Information retrieval. Arxiv, OpenAlex, PubMed, PubChem, ChEMBL, ClinicalTrials, USPTO, FRED, arXiv scanner, grounded citations, and the research-lookup and parallel-web skills that compose them.

Hobby

—

Mineral collecting, crystal chemistry, stargazing, hiking, foraging, and the print-media and mineral-collecting skills that back the hobby sites. The empire's editorial depth comes from the time invested here.
03

Twelve skills that earn the daily rotation

If you load twelve of the 424, load these. They are the ones that turn into durable artifacts inside seven days of use.

📚

Legion

scholar-omega — research director

The continuous-intelligence skill. Loads the agent's persistent memory, then runs a daily or weekly scan across frontier AI sources, biotech, crypto, and empire-relevant opportunities. Produces an intel brief with sources and a recommended action. Use it when you need to know what changed in a domain without spending the hours yourself.

# invoke
python "C:/Users/BRC/AppData/Local/hermes/scripts/legion_agent_memory.py" bootstrap --agent scholar-omega
python "C:/Users/BRC/AppData/Local/hermes/scripts/legion_agent_memory.py" recall --agent scholar-omega "frontier AI agents August 2026"
# deliverable: 200-400 word brief with cited sources → #competitor-intel
Owner 👑 PrimeCadence daily / weeklyChannel #competitor-intel
🛠️

Empire

empire-pulse — heartbeat watcher

Single-line skill: probe all canonical empire services, all cron jobs, all public sites, count free RAM, and emit a single JSON object. The Monitor-Epsilon cron reads that JSON every 5 minutes. Without this skill, you learn that your empire is down from a customer.

# invoke
python "C:/Users/BRC/AppData/Local/hermes/scripts/empire_pulse.py"
# delivers to Discord #empire-infra via discord_pulse_render.py
Owner 📡 MonitorCadence 5 minChannel #empire-infra
📋

Productivity

morning-standup — daily empire brief

Aggregates the pulse, the BMAD scorecard, the wallet snapshot, the kanban active count, the last 7 days of legion work, and the top 3 grow.md actions into a single one-page markdown brief. Delivered to Telegram at 8 AM ET, written to ~/.hermes/morning-brief.md for the day. Reads-only, fail-soft — never blocks the brief if one source is down.

# invoke (manual or via cron job 1f8105185a53)
python "C:/Users/BRC/AppData/Local/hermes/scripts/empire_summary.py"
type "C:/Users/BRC/.hermes/morning-brief.md"
Owner 👑 PrimeCadence daily 08:00 ETChannel Telegram + file
🔒

Security

security-headers-rollout — empire-wide OWASP baseline

Lifts every FastAPI service from F to A on the local-headers matrix in one batch deploy. Ships the six OWASP-recommended headers (HSTS, CSP, X-Frame-Options, X-Content-Type-Options, Referrer-Policy, Permissions-Policy) plus COOP/CORP. Verified 42/42 across 7 services on rollout day.

# in any new FastAPI server
import sys
sys.path.insert(0, r"C:/Users/BRC/AppData/Local/hermes/scripts")
from empire_security_headers import SecurityHeadersMiddleware
app = FastAPI(...)
app.add_middleware(SecurityHeadersMiddleware)  # immediately after app =
Owner 🛡️ GuardianCadence per new serviceCoverage 42/42 services
🦙

AI / ML

llama-cpp — local 35B inference on Bert's AMD box

The local-LLM skill. Runs the 35B Ornith model with --ctx-size 32768 on the RX 9070, exposes OpenAI-compatible /v1/chat/completions on :8084, and is the inference path for every empire agent that needs an LLM. Free, fast, and never sends a customer prompt to a third party.

# invoke
llama-server -m "C:/models/ornith-35b-q4.gguf" -c 32768 --port 8084 -ngl 99
curl http://127.0.0.1:8084/v1/models
curl http://127.0.0.1:8084/v1/chat/completions -d '{"model":"ornith-35b","messages":[{"role":"user","content":"hi"}]}'
Owner 🦙 PrimePort :8084Cost $0 / 1k tokens
🔧

Business

be-repaired-intake — handyman lead capture

Parses incoming messages from the BE Repaired phone, the website contact form, and the Facebook page into a single lead object with name, address, requested service, and urgency. Posts to the leads API on :8091 and to a Discord channel. Catches the "I left a voicemail" lead before it ages out.

# invoke
python "C:/Users/BRC/AppData/Local/hermes/scripts/be_lead_pipeline.py" --source sms
python "C:/Users/BRC/AppData/Local/hermes/scripts/be_repaired_supervisor.py" status
Owner 🔧 FixerLead time < 5 minChannel Discord #be-leads
🔍

Research

parallel-web — deep multi-source research

The "I need to know the actual answer" skill. Fans out a query across 5-7 search backends in parallel, deduplicates results, and emits a single structured brief with sources, claim-level citations, and a confidence flag. Use when a single web_search is not enough and the answer has to be defensible.

# invoke
from hermes_tools import web_search, web_extract
results = web_search("smallest viable local LLM August 2026", limit=10)
# then web_extract the top 5 URLs, compose the brief in 600 words
Owner 📚 Scholar-OmegaBackends 5+ parallelCost free / API
🪨

Hobby

mineral-collecting — Avery County field guide

Curated locality data, KML packs, and a public field journal for the Spruce Pine pegmatite district, the Mitchell County garnet corridor, and the McKinney mine kyanite dumps. Backed by real data at ~/Desktop/Mineral-Collecting/ and shipped as minerals.b3rt.dev. The reason the empire's editorial voice has actual weight.

# invoke (read-only)
ls "C:/Users/BRC/Desktop/Mineral-Collecting/"
curl https://minerals.b3rt.dev
# KML pack: see empire/sites/minerals.kml
Owner 🪨 PicassoCadence per field tripChannel minerals.b3rt.dev
📜

Empire

sitemap — auto-generated discovery file

Walks the SITES tuple, the hub's static page allowlist, and the public b3rt.dev subdomains to produce a valid sitemaps.org XML file with per-surface priority and changefreq. Served at /sitemap.xml from the hub. Pair with robots.txt and you have machine-discoverable empire.

# invoke
python "C:/Users/BRC/AppData/Local/hermes/scripts/empire_sitemap.py"
curl http://127.0.0.1:8098/sitemap.xml | head -20
Owner 📚 ScribeCadence per SITES editOutput 4181 bytes XML
🤖

Empire

subagent-launch — read this first

The discipline layer. Before any subagent is spawned, this skill is read. It enforces: the task has a hand-back contract, the workspace is scoped, the budget is bounded, the verification step is named, and the parent keeps a checkpoint. Subagents without a contract are noise.

# in any subagent prompt
## CONTRACT
Input:    <what the subagent gets>
Output:   <the hand-back format>
Budget:   <max tool calls, max wall time>
Verify:   <how the parent checks the work>
Then:     <what to do on success vs. on partial>
Owner 👑 PrimeCadence per subagentFailure cost hours lost
🌅

Productivity

imperator-brief — daily 6-section brief

The closest the empire has to a chief-of-staff report. Pulls the morning brief, the BMAD scorecard, the empire grade, the wallet, the pulse, the kanban, the cron summary, the drift detector, the recent wins, and composes a single 6-section markdown brief tagged for the day. Delivered on demand to the Discord channel of choice.

# invoke
python "C:/Users/BRC/AppData/Local/hermes/scripts/empire_summary.py" --save
type "C:/Users/BRC/.hermes/morning-brief.md"
Owner 👑 PrimeCadence on demandLength ~1200 words
🎓

Productivity

spec-driven — start features with SPEC.md

The non-negotiable first step. Before any feature, bug fix, or build task, the spec-driven discipline writes a SPEC.md with the four questions: what entities, what columns/states, single-user or multi, what integrations. The empire's entire build catalog follows this — every new tool has a SPEC.md that survives the build.

# invoke (workflow)
1. Write SPEC.md with the four questions
2. Pick the stack (FastAPI + SQLite + vanilla JS for tools)
3. Pick a free port (verify with netstat)
4. Build → health check → create/read/edit/delete → screenshot
Owner 👑 PrimeCadence per buildFailure cost rework
04

The full table at a glance

Every category, a representative sample. The library is bigger than this table, but the shape is the shape.

CategoryCountRepresentative skillsOwner
Legion20+scholar-omega, builder-sigma, cypher, fixer, guardian, bard, maverick, sparks, scribe, warhol👑 Prime
Empire40+empire-pulse, empire-phalanx, empire-skill-eval, empire-archive, empire-build-burner, empire-os-patterns, hub-ops, sitemap, supervisor, sitemap📡 Monitor
Productivity35+morning-standup, kanban-triage, handoff, error-handler, spoken-empire-briefing, document-to-action-items✍️ Bard
Security15+security-headers-rollout, dependency-audit, secret-scan, mcp-oauth-setup, tailscale-remote-access🛡️ Guardian
AI / ML60+llama-cpp, local-llm-amd, transformers, segment-anything-model, tts, comfyui, huggingface-hub, weights-and-biases🦙 Prime
Business30+be-repaired-intake, be-repaired-quote-builder, be-jobs-pulse, sparkling-bookings, be-repaired-supervisor, lead-enrichment, welcome-sequence-automation🔧 Fixer / ✨ Sparks
Research80+arxiv, openalex-database, pubmed-database, clinicaltrials-database, fda-database, uspto-database, parallel-web, grounded-citations📚 Scholar-Omega
Hobby25+mineral-collecting, on-this-day, songwriting-and-ai-music, popular-web-designs, design-md, excalidraw, manim-video, p5js🎨 Warhol / 🪨 Picasso
Hack / infra50+recursive-build-orchestrator, recursive-empire-build, empire-meta-loop, empire-build-burner, finn-loop-usage, project-lifecycle, kanban-triage👑 Prime
Software dev40+spec-driven, simplify-code, test-driven-development, requesting-code-review, pr-review-fast, systematic-debugging, spike, plan🛠️ Builder-Sigma
Creative20+claude-design, sketch, baoyu-infographic, architecture-diagram, ascii-art, songsee, comfyui, on-this-day🎨 Warhol
Other / utility30+windows-cron-no-popup, windows-locked-pid-recovery, computer-use, hermes-web-vision, hermes-state-migration, empire-backup, disk-cleanup, finn-loop🔧 Operator
05

How to invoke a skill

There is no central dispatcher. The pattern is load the SKILL.md, follow the procedure, verify the artifact.

A skill is a markdown file. To load it, point an agent or an operator at the path. The agent reads the frontmatter (name, description, tags, version, owner) and the body (trigger, procedure, references, verification, pitfalls). The agent then follows the procedure in its own context. When the procedure produces an artifact, the agent writes the artifact, returns a hand-back summary, and records what worked and what surprised it.

There is no need for a registry, a manifest, or a search index. The file system is the index. If you want a list of all skills, run find ~/AppData/Local/hermes/skills -name SKILL.md or hit http://127.0.0.1:8098/api/skills for the cached JSON manifest. The manifest is regenerated every 60 seconds and never lies about what is on disk.

To write a new skill: notice a procedure that has repeated three times. Open a new SKILL.md file. Write the frontmatter (name, description in < 60 chars, tags, version 0.1.0, owner). Write the body (when to use, the procedure, references, verification, pitfalls). Commit it. Tell one other operator or agent it exists. Watch whether it gets used a fourth time. If yes, you have a skill. If no, you have a draft. Both are useful.