Sending the OPL newsletter from Telegram
I connected MailerLite’s API to our OPL Hermes assistant so I can write and send newsletters from Telegram. On 18 September, I tested it and then used it for our full mailing list.
Open noteThe notebook
Tested, used, written down.
Practical notes from my AI work: the setup, what happened when I tested it and what I took from it. New ideas are labelled separately.
Collected in September 2026. Earlier experiments are described retrospectively.
From an idea to a website I can maintain
I’ve learned to build websites with Codex and Claude Code, keep the source in GitHub and publish through Cloudflare Pages. The useful part is being able to come back and change them.
Open noteWhen memory needed a test, not another promise
I keep durable notes in Obsidian and use Hermes memory to help retrieve useful facts. A September fault made me check what was actually being saved and found.
Open noteFinding the right way to use Hermes
I’ve tried Telegram, Matrix and Discord for chatting with Hermes, alongside Hermes Desktop. The Dashboard has a different job: it’s where I manually edit settings such as model routing.
Open noteRunning two assistants on one Raspberry Pi
I rebuilt Neebs and Spud on a Raspberry Pi, then tested Telegram access, file boundaries and approval before sending email.
Open noteReading text from images with Ollama
I liked GLM-OCR when I tried it through Ollama. Looking at a simpler LM Studio setup then raised a different question: would the same model work in the tool I wanted to use?
Open noteLocal models with LM Studio and AnythingLLM
I connected LM Studio and AnythingLLM on my M1 Max to explore local models and working with my own documents. The practical limits were as useful to learn as the possibilities.
Open noteAn early experiment with an Ableton teaching assistant
I built a custom GPT around Ableton manuals and course material so students could ask questions about what they were learning.
Open notePicking up the same project with another AI tool
Moving Terrace Archive from Claude Code to Codex showed me the value of keeping the project somewhere I controlled.
Open noteGiving Spud a place in the OPL conversation
We use Spud from our OPL Telegram room for the admin around the band. Giving it a place in a shared conversation meant deciding when it should join in and what information belonged there.
Open noteWhy I keep coming back to local AI
I’m interested in having more control over the tools I use. Exploring local AI has also made me more careful about what ‘local’ actually means.
Open noteWhat I want to try with video editing
I’ve made videos before. My next question is where computer use could help with the editing work around them.
Open noteMaking room for TouchDesigner experiments
I want to explore help with TouchDesigner and visual workflows, then document what the experiments actually demonstrate.
Open noteTrying a local chat interface with Open WebUI
Open WebUI was one of the interfaces I tried around Ollama. The setup notes are also a useful reminder that an installed app and a complete workflow are different milestones.
Open noteThe morning report, and knowing when to stop adding more
I built a morning report that brought useful information together in Obsidian. I also learned that adding more scheduled messages could make the system harder to live with.
Open noteThe AI videos I want to make
Two tutorials a month, built around tasks I’ve actually tried. I want to show the useful parts, the corrections and what still needs a person.
Open noteThinking through local AI for sensitive documents
My local-model experiments made me think about businesses that want useful answers from their documents without sending everything to a cloud model.
Open noteThe next computer-use test: my collection
I want to use ChatGPT with computer use to help prepare the sale of my football programmes and NME magazines, building on the work behind Terrace Archive.
Open note