Week 6 of 6

Agentic Creativity

Voice, Plugins, Local Models, and Show and Tell
Live on Zoom · 2.00–4.30pm ET
27 September 2026
Derrick Schultz

Six weeks

Week 1 · Aug 16

Introduction to agentic creativity and Hermes Agent setup

Week 2 · Aug 23

Models, text generation, and everyday tasks

Week 3 · Aug 30

Skills, MCPs, and visual capabilities

Week 4 · Sep 13

Long-running agents and how they change over time

Week 5 · Sep 20

An introduction to agentic coding

Week 6 · Sep 27

The last hour of material, then show and tell

Week 1 we installed a toddler. Today: three things I owe you from earlier weeks, then we find out who it grew into.

Today’s schedule

2.00

Voice: talking to your agent instead of typing

2.15

Plugins: what they add that skills and MCPs don’t

2.25

Local models: free, private, slower, dumber

2.45

Monday morning: keep it, pause it, or stop it

3.00

Break

3.05

Show and tell, ten people

4.10

The class agent presents, then goodbyes

today’s question

What did you end up with, and how much of it is yours?

You answer at 3.05. The first hour is the three topics I promised and didn’t get to.

Voice

2.00–2.15 · 15 minutes

from the week 2 discussion

You’ve typed to your agent for six weeks. You don’t have to.

Someone said in Week 2 they’d got into the habit of “word-vomiting” at their agent instead of writing prompts. That’s a real technique. Here’s how to do it in Hermes.

/voice

Push to talk, or say “hey Hermes.” Free, and it runs on your machine.

In the terminal or desktop

/voice on
Ctrl+B to record. Stops after 3s of silence.

/voice tts it talks back, sentence by sentence
/voice status

# hands-free
/wake on "hey hermes" · detection stays on-device

On your phone

# Telegram or Discord
send a voice note → it transcribes and answers
/voice on replies as voice when you spoke
/voice tts replies as voice always

# what it costs
speech-to-text: local whisper, free, ~150MB
text-to-speech: Edge TTS, free
ElevenLabs if you want a voice you'd put in a film

Speaking mid-reply cuts it off and sends what you said. It’s the closest these tools get to a conversation. · Hermes: voice mode

Talking produces a different kind of prompt. Use that.

Good for

Briefs. Describing a feeling, a reference, what’s wrong with an image. You say more, and more of it is true.

Bad for

Paths, commands, anything with a number in it. Type those.

The habit

Talk for a minute, then: “Restate that as a brief before you do anything.” Fix the restatement, not the ramble.

Where it pays off

The Week 4 BRIEF.md. Dictate it on a walk. Let the agent write the file. Read it once at your desk.

Plugins

2.15–2.25 · 10 minutes

Skills teach it. MCPs connect it. Plugins change what Hermes itself can do.

A skill

Instructions. A method the agent reads when the task matches. Week 3.

An MCP

A bridge to a service that already exists. TITLES, Notion, Google. Week 3.

A plugin

Code inside Hermes. New tools, new slash commands, a new memory backend, a new messaging platform, a hook that runs after every tool call.

Who writes them

Programmers, mostly. You install them. Honcho from Week 4 is one. So is every messaging platform you connected in Week 1.

You’ve been using plugins since Week 1 without the word. · Hermes: plugins

opt-in, every one

Nothing loads until you enable it. Three worth enabling today.

Built in, off by default

disk-cleanup
deletes the temp files a session leaves behind

security-guidance
warns when it writes code that looks dangerous

hermes-achievements
badges from your session history. useless. fun.

# others: spotify, google_meet, image_gen/openai

Commands

hermes plugins list
hermes plugins enable disk-cleanup

# from the community index
hermes plugins search media
hermes plugins install owner/repo
scanned on install: safe / caution / dangerous

hermes plugins capabilities <name>
what it asked for vs what you granted

Same rule as MCPs in Week 3: a plugin is a permission. Read the scan verdict, and don’t install one that wants tools.override unless you know why.

Local models

2.25–2.45 · 20 minutes

the week 1 promise

A model on your own laptop is free, private, slower, and dumber. Sometimes that’s the right trade.

Nothing leaves the machine. No bill. No provider deciding to retire your model. And a noticeably worse agent.

live demo · apple silicon

Twenty minutes to a local agent. Most of it is the download.

The model server

# llama.cpp, the standard
brew install llama.cpp huggingface-cli

huggingface-cli download unsloth/Qwen3.5-9B-GGUF \
Qwen3.5-9B-Q4_K_M.gguf --local-dir ~/models
~5GB. or the current small Qwen.

llama-server -m ~/models/Qwen3.5-9B-Q4_K_M.gguf \
-ngl 99 -c 131072 --cache-type-k q4_0 --cache-type-v q4_0
serves at localhost:8080

Point Hermes at it

hermes model
  → Custom endpoint
  → http://localhost:8080
  → model name from the server

# the easier path, with a GUI
omlx from omlx.ai  Apple's MLX, faster generation
  → serves at 127.0.0.1:8000

# first turn silent for a minute? that's prefill, not a hang.

Hermes needs at least 64K of context, and sends its whole system prompt every turn. On a slow machine that’s the wait. · Hermes: local LLMs on Mac

Quantization is why a model fits in your laptop at all.

What it is

Rounding the model’s numbers to fewer bits. Q4 means four. About a quarter the size, a little dumber. The context cache can be quantized too.

8–16 GB Mac

A 9B model at Q4. About 5GB on disk, 10GB in memory with a quantized cache.

32 GB and up

27B to 35B models. Noticeably better. Still not a frontier model.

Speed

Fine on an M-series chip. Painful on Intel. Every model file is gigabytes; budget an evening for the download.

When a local model is fine, and when it isn’t.

Fine

The Week 2 digest. Renaming and sorting. First-pass research triage. Anything you’d check anyway.

Fine with skills

Week 1’s claim holds: a small model with good skills does most rung-one work. This is where that pays off.

Not fine

Vision, taste, long tool chains, the class agent. Keep a cloud model as the auxiliary for those.

The real reason

Privacy. Tax returns, client work, unreleased pieces. A local model is the only one that never sees a server.

Hermes can mix them: local as the main model, a cloud model for vision and the goal judge. Week 3’s two-models-two-jobs slide, again.

Monday morning

2.45–3.00 · 15 minutes

Decide on purpose. Don’t let it just keep running.

Keep it

Pin the model. Read the output folder weekly. Rewrite the brief when your work changes, not when the output gets boring.

Pause it

hermes cron pause <job>. Nothing is lost. The log and memory wait for you.

Stop it

hermes cron remove <job>. Then export the sessions you want to keep: hermes sessions export. Delete the cloud instance if you made one. It bills while stopped.

Either way

Turn off any key you’re not using. Check the provider dashboard once more in October.

The most expensive agent is the one you forgot about. Week 1’s cost slide is still true.

Where to go from here, by rung.

Rung one · clear the desk

Keep the digest. Add one more admin job a month. This alone pays for the class.

Rung two · works next to you

Give it your real project folder and a real brief. Use the critique skill on your own work once a week.

Rung three · works without you

Keep one long-running job alive for three months. Read the log monthly. Decide in December whether any of it is yours.

Any rung

Build one size-1 tool every time you catch yourself doing something tedious twice.

Break

Stand up. Refill. Open what you’re going to show.

3.00–3.05 · 5 minutes

Show and tell

3.05–4.10 · 65 minutes · ten people

Six minutes each. Three things, two questions.

One thing it made

Screen share. From the job, the tool, or both. The output, not the setup.

One correction

A line you changed in a brief, a SOUL.md, a README, or memory, and what came out differently after.

One refusal

The judgment you kept for yourself. Has it moved since Week 1?

Two questions

From the room. Evidence before interpretation. I keep time.

Anyone who’d rather not present can show one image and take one question.

the protocol we used on the agent, now for each other

Evidence. Inference. Then advice, if asked.

WHEN SOMEONE SHOWS WORK

1. Say what you see. One concrete thing.
2. Say what it might mean. Label it as a guess.
3. Ask a question before you offer a change.
4. If they ask for a next experiment, offer one. Small.

NOBODY SAYS WHETHER IT IS FINISHED. INCLUDING ME.

The agent has been held to this for four weeks. Fair’s fair.