AI Personal Assistant Playbook: Use Hermes/Xena Deliberately

What an assistant is best at

A capable assistant is not a replacement for judgement. It is leverage for research, synthesis, drafting, repetitive operations, monitoring, and implementation—with verification proportional to impact.

Task	Good delegation	Keep human judgement
Research	Compare sources, build a briefing	Decide what to believe/do
Writing	Draft, restructure, tailor	Claims, tone, final external commitments
Home lab	Inspect, diagnose, prepare runbooks	Destructive/recovery decisions
Business	Prepare lists, drafts, analyses	Targeting, outreach approval, promises
Finance	Categorise, summarise, remind	Transfers, tax decisions, investment decisions

The request pattern that gets reliable outcomes

Outcome: what finished looks like
Context: stack, audience, constraints
Scope: what may be changed
Quality bar: build/test/source requirements
Safety: what must not be exposed or automated

Example: “Build a KB guide on X, use the existing visual system, run the site build, browser-test the route, deploy it, and do not publish secrets.”

Tools: choose the right execution surface

Need	Best approach
Current facts	Web/search or direct source inspection
Local files/code	Read, edit, build, test with real outputs
Home Assistant state/control	HA tools/API with scoped services
Repeated timed task	Cron job with a self-contained prompt
External app workflow	Connected/app-specific tools, with active auth
Creative media	Image/audio/video tools with clear prompt and review

A tool result is evidence; a fluent description is not. For important tasks, ask for build logs, links, IDs, or other verifiable artifacts.

Skills, memory, and projects

Skills store reusable procedures—publishing the KB, operating HA APIs, coding workflows.
Memory stores stable preferences and environment facts—not temporary task logs or secrets.
Project trackers store multi-step programme state in version control.
Cron jobs make a recurring action durable; their prompts must stand alone because they run without this conversation.

Model selection

Choose for the task, not prestige:

Work	Preference
Simple classification, status, mechanical writing	Low-cost fast model
Complex code/reasoning/review	Stronger reasoning model
Sensitive or private material	Use only approved local/private configuration; minimise data
Long research	Model plus source links and staged synthesis

Always budget for verification: a cheaper model that requires five repairs can cost more than a stronger one used once.

Voice interaction

Voice is excellent for capture, quick control, and conversation. For actions that require precision—names, IDs, money, destructive commands—confirm the interpreted result before execution. Treat transcription variants of “Xena” as likely speech-to-text drift, not a change of assistant identity.

Privacy and safety boundaries

Never expose tokens, passwords, database dumps, private locations, or personal financial records in a generated artifact.
Use least privilege for integrations and revoke unused access.
Prefer monitoring and approval gates over automatic remediation for consequential systems.
Do not turn a draft into external outreach, a transaction, or a system change without the authorised scope.

A healthy operating cadence

Capture ideas quickly → turn worthy work into a project/guide → verify delivered artifacts → retain reusable process → prune stale automation

The assistant should make the system simpler over time. If a workflow creates noisy notifications, opaque complexity, or a dependence on forgotten context, redesign or retire it.