Hermes Skills: Reusable, Approval-Gated Workflows for a Private AI Assistant

Hermes skills are reusable procedures for a private AI assistant. The decision rule: if a task repeats, has clear inputs, and should produce the same kind of output every time — lead reply drafts, meeting briefs, support responses, weekly reports — it belongs in a skill. If you would otherwise retype the same instructions in chat each week, that is a skill waiting to be written.

A skill is not a magic plugin and it should not grant broad authority. A good skill tells Hermes when to run, what to inspect, which tools are allowed, which actions are blocked, what to produce, and when to ask the owner for approval. For a founder, consultant, agency, or tiny team, skills turn "do this the way I like it" into an auditable operating procedure.

If you are new to the broader setup, pair this guide with Private AI Executive Assistant Setup and Approval-Gated AI Agents — skills are where those approval gates get enforced in practice.

What a Hermes skill does

A skill answers one practical question: how should Hermes do this kind of work?

A strong skill includes:

• purpose: the result the workflow should produce;
• trigger: when Hermes should use it;
• inputs: messages, files, systems, or context required;
• allowed tools: what the assistant may read or draft with;
• blocked actions: what it must not do;
• approval gates: when the assistant must stop and ask;
• output format: what the owner receives for review;
• receipt: what gets logged after the run.

The skill should be short enough for an owner to inspect. If it reads like a policy binder, split it.

Skills vs prompts vs automation

These three get confused.

A prompt is a one-off instruction. It is useful for experimenting, but fragile when the same task repeats.

A skill is a reusable workflow. It standardizes the steps, output, tools, and stop conditions.

Automation is when that workflow runs from a trigger or schedule. Automation should come after the skill has been tested, not before.

A safe progression is: prompt to explore, skill to standardize, scheduled or triggered workflow to scale. Do not skip the middle step.

Good skill candidates

Create a skill when the task repeats and mistakes are easy to define.

Good first candidates:

• lead follow-up draft from a form or email;
• meeting brief from calendar, CRM, and recent notes;
• support ticket summary and response draft;
• daily owner briefing from open loops;
• weekly project status report;
• content outline review against brand rules;
• invoice follow-up draft;
• research brief with source links.

Weak candidates:

• "handle sales";
• "be strategic";
• "run support";
• one-off research;
• anything where the owner cannot name the stop condition.

The best first skill saves review time without pretending the assistant owns the decision.

Example: lead follow-up skill

A tiny agency might use a lead follow-up skill like this.

Purpose: prepare fast, accurate reply drafts for qualified inbound leads.

Use when: a new lead arrives from an approved inbox, form, referral message, or CRM queue.

Allowed actions: classify the lead, summarize context, draft a reply, suggest a CRM note, and ask for approval.

Blocked actions: send email, quote price, promise availability, mark a deal closed, delete the source message, or change CRM stage without approval.

Approval packet: source message, lead classification, proposed reply, CRM note, risk level, and exact action requested.

A real run: a referral lead emails in on a Tuesday afternoon. The skill fires, reads the message, pulls the last two interactions from the CRM, and classifies the lead as "warm referral, ready to book." It drafts a reply that acknowledges the referral by name, offers two time windows from your calendar, and stops before quoting price. You get the approval packet in one read, approve, and the assistant sends the reply, updates the CRM to "Contacted," and logs the receipt to memory. Total owner time: under a minute. Without the skill, you would have done the same five steps manually and probably waited until the next morning.

That is specific enough to help and narrow enough to review.

Put approval gates inside the skill

Approval rules should be explicit, not implied. Require owner approval before Hermes:

• sends customer messages;
• changes CRM status;
• posts public content;
• edits production systems;
• deletes or archives source material;
• quotes price, timeline, discounts, or refunds;
• makes legal, financial, medical, hiring, or reputation-sensitive commitments.

For each approval, tell Hermes what evidence to show. The owner should see the source, proposed action, exact output, risk, and where the receipt will be logged.

A bare "approve?" message is too thin. The point of the skill is to make approval fast because the decision packet is complete.

Skills, memory, and runbooks are different

Do not use skills as a dumping ground for every business fact.

Use this split:

• Runbook: the business operating rules for a workflow.
• Skill: the reusable procedure for one task inside that workflow.
• Memory: stable preferences and durable facts that should persist.
• Project file: current state the assistant can reread.
• Chat: the immediate request and output constraints.

Example: "never quote price without owner approval" belongs in the runbook and relevant skills. "The owner prefers short customer replies" can live in memory. "This week is booked" belongs in current status, not permanent memory.

Test before trusting

Run the skill against real but low-risk examples. Use at least three cases:

1. normal case;
2. missing-context case;
3. risky case involving customer trust, money, timeline, or policy.

Check whether Hermes used the right source, followed the right order, stopped at approval, produced the expected format, and flagged uncertainty instead of guessing.

If the same correction appears twice, update the skill. Do not keep retyping the same instruction in chat.

Keep skills narrow

A skill should not be "Sales Assistant." It should be "Lead Reply Draft," "Proposal Follow-Up Checklist," or "CRM Note Preparation."

A useful rule: if a skill needs more than one owner, more than one approval model, or more than one source of truth, split it.

Narrow skills are also easier to retire. Old skills become hidden bugs when offers, tools, credentials, or approval policies change.

First-week setup

1. Choose one repeated low-risk workflow.
2. Write the current manual process in five to ten bullets.
3. Add allowed tools and blocked actions.
4. Add approval gates and the exact review packet.
5. Define the output format.
6. Test on three examples.
7. Save corrections into the skill.
8. Use it for a week before expanding scope.

The first skill should save review time, not prove autonomy.

Recap

Hermes skills are reusable, reviewable operating procedures for a private AI assistant. Use them for repeated work with clear inputs, outputs, tools, and approval points. Keep them narrow, test them with real examples, separate them from memory and runbooks, and retire stale skills quickly.

Next step

For the broader setup pattern, read What Is a Private AI Executive Assistant? and Approval-Gated AI Agents. If you want a done-for-you implementation path, see Private AI Executive Assistant Setup.