Ultras & how my AI usage evolved | 1H 2026
1H 2026 - Done! WHAT.
Hello there!
In the midst of watching the glory of World Cup, we Americans are about to celebrate the USA’s 250th birthday.
Today I’ll show you a clip from my running, as well as a short “How my AI usage evolved” piece.
Ciao!
Updates from me:
Angeles Crest 60k Ultra Training run: I recently completed a 60k training run, making it the second ultra I’ve ever done. I have a 25 second clip about this later on.
Leadville Silver Rush 50: I have my next racing ultra on July 11, a 50 Miler. Unfortunately, that is in danger of postponement due to the Willow Fire nearby, and I won’t know until Tuesday on what the call is from the race organizers as the Willow fire isn’t contained. It’s a logistic challenge as I fly out for altitude acclimation July 3. I suspect it’ll be cancelled, but we shall see.
KEEP CALM AND CLAUDE ON: I am currently experiencing a renaissance in my claude usage across the board, and I have crossed over into the line of prejudice against those who are “skeptical” of AI, and I’m an absolute madlad about it. Why into the line of prejudice? Because practitioners of accounting, tax, and finance that I’ve encountered disappoint me on their attitude and usage in the same energy as those who dismissed digital transformation. I expand on this later.
Oh we so back, EVE Frontier Cycle 6 is here and there is a 5 day free trial for you to get in. If you are interested and can commit 90 minutes for your first run, lets play! DM me. The game is starting to take on its own identity, and it will be a ruthlessly difficult hardcore space driven mmorpg that’d make Isaac Asimov twinkle.
Angeles Crest 60k Ultra Training run
I’m at the point that I require minimal heads up in doing impatient hiking ultras in the forest.
Here’s a 25 sec video about that:
I run Mount to Coast, I wear Janji, and I do a lot of DIY Gels. You can follow me on Strava here.
How my AI usage evolved
Six levels, traced from the actual project folders.
Up until 12 weeks ago, I was an avid ChatGpt Pro user. My overall competency with AI Tooling has significantly increased as a result of:
my need to assess AI SaaS Vendor tools so I can tell the counterparty its junk, and
my need to control and reduce a specific variable: shadow debt, which in this context I am referring to unsanctioned or ungoverned AI Tool usage as well as beating “ai for the sake of using ai”.
That changed when my wife showed me a few projects she made in Claude, and it was from that exposure I’ve had my own industrial revolution. Several times.
Level 1: prompt and deliver
One-off prompts, single session, file out.
Ask a question, get a deliverable back.
Everyone starts here, and I was here for like 12 months on ChatGPT.
Here I just saw LLMs as “search engines but faster”.
Using AI this way made me skeptical about AI adoption and I had fallen for the mass opinions on hallucinations and poor performance. I imagine a lot of people who are exposed to AI Tooling through less than great tools also see lacklusterness.
That said I was able to burn a lot of tokens in my first month and made a game out of it and seperately, participated in a hackathon, so that was cool.
Level 2: contracts and boundaries
The first time I wrote rules for the AI instead of requests.
A read-only “observer contract” spelled out what the agent could and couldn’t touch, and a handoff folder held the artifacts (outputs) of those guardrails.
Here I saw LLMs as “search engines, but using advanced search”
Here I started to do those seed-doc / personal preferences loading. You’ll see those posts about “max token usage” and so forth which is a very stupid thing to do, and I did it too. People go crazy about their seed docs, and many people still do.
“What is your purpose?” You spread butter.
Level 3: session handoffs
Written handoff specs so one session’s context survived into the next.
The work became versioned and continuous instead of restarting from zero every chat and having to rebuild momentum.
I am a hardcore procrastinator and so the moment I perceive exceptional momentum to “get my flywheel started” i.e. I face friction, I’m just not going to do it and instead I will defer it as I require a lot of energy to overcome my inertia.
This is almost like a seed or profile doc, but on a per project basis. Lots of heavy context had.
This came about actually because of my need to collaborate met my need to want to take a vacation and “return to the project” later on. Handoff documents were everything to me! They are like saved states for video games.
Level 4: engine building
I turned my own workflow into reusable templates: role-specific priming prompts, a live artifact manifest, cheap models for grunt work and expensive ones for judgment calls, token budgets per role, and a rule that corrections flow upstream to the file/chat that owns the fact.
Effectively, I created an assembly process or a kitchen and any kinks, I’d know exactly where the bottleneck is and how to troubleshoot it. In case you don’t know, I play a lot of Factorio so logically I am capable of thinking this way.
This was a significant shift, from using AI as a glorified search engine, to a factory of systems with me standing as an operator.
Oh fun fact, this is where I deleted all seed and context docs, and anything I consider a “vain fact” that only served to “make me look big like a peacock”. I don’t even say thank you anymore because that costs tokens.
Level 5: parallel orchestration
Several AI roles running at once under one conductor, sharing assumption files and citing sources. Effectively I discovered in manufacturing of digital data how to operate multiple AI Agents and use another agent for orchestration and conducting, and then made that AI Agent hot swappable so I can spin up any agent to check in on my agents.
This was also the moment I allowed for significantly more complex outputs to go out to people under the expectations I am making, or preventing, decisions based on what was made.
Level 6: meta-orchestration
A homepage that scans all my projects and launches headless AI build sessions when I click a tile, streaming the output live. AI that launches AI. My goal was this: “Create a system that prompts itself”, wherein “System” is a methodology. But I took it literally and made a game engine. This is dangerously fun.
I have a fun browser video game and card game that represents this…and it does execute and create things using Claude while I’m in those games. It’s hilariously wasteful, I love it.
What I learned
There was a focus on prompts. I’m terrible at it and now have my Claude create the prompts for us to keep going. What I’m great at is process engineering and orchestration, and making assembly lines so I can see where I can push the limits.
That was fun.
Prompting was innately a skill that did not “compound” for me. Process engineering was: manifests, correction flows, role boundaries, and doing so it allowed me to go from Level 3 to 5 pretty fast.
When I actually go to use technology, I’m superstitious enough and skeptical enough to know where it will fall short or “hallucinate”, that I can effectively eliminate those issues by design. I’m proud of that.
I have AI outputs auditable and traceable, build processes that allow me to correct and detect for errors on the fly. For the longest time I plateaued at level 1 or 2, and then I was able to jump past that by learning AI Orchestration. Now I teach that, and it’s allowed me to stop asking the AI to do tasks and started building the environment it operates in so I can go do something else.