My AI journey has two starting dates. The first is 2023, when I began using it to write social media copy. The second is March 2026, when I started working with Claude Code and, for a whole month, got slower at my job.
Copy that looked finished
In 2023 I was a junior account manager at Iridian, in Bogotá, running content calendars with a small team of designers, media buyers and community managers. AI had just arrived at work, and I used it the obvious way: to write the copy for our clients' posts.
It was fast, and it was slop. Nobody called it that yet; the term caught on later. The problem was already there, though. The captions were correct, on topic and interchangeable. You could swap the brand name and post them for someone else.
That was my first lesson, learned before I had a word for it: a text can be finished and still say nothing about the brand behind it.
GPTs, projects and too much context
Later that year I moved to Boston for a master's at Hult and started building custom GPTs: assistants with their own instructions and my files attached. I gave them everything I had, PDFs and Word documents, one after another. The more I gave them, the worse they got. The answers changed from one day to the next and I couldn't tell why.
Then came projects, and a habit that stuck: a new chat for each task, with only what that task needs. I didn't know it yet, but I was managing context, the amount of information a model can hold before it loses track of what matters.
- Early 2023. Iridian. AI writes the copy for our posts: faster, and slop.
- Late 2023. Hult. Custom GPTs with too many files, and the answers get worse.
- 2024. Projects, and a new chat for each task.
- 2025. First automations.
March 2026: slower before faster
In 2025 I experimented with automations. The real start came in March 2026, with a challenge at work. The company decided to focus on middleware, and I was given the job of training someone in Boomi and MuleSoft, along with the marketing department's processes, for an SDR role. It was the company's first SDR and its first formal outbound strategy.
So I designed, with my agents, a system that could train the SDR and measure the progress at the same time. It ended up as a mix of Excel and agentic workflows, where the SDR logged things like the people who accepted each connection request.
That was when I started working in Claude Code and Codex, in March and April, and when the concepts started to make sense. Tokens came first: every file an agent reads has a cost, and the budget runs out. I moved to Claude's $100 plan.
My productivity dropped. I was doing my job and, at the same time, learning how to hand parts of it over, and the second part took most of my attention.
The tools are fast. Learning to work with them takes time.
That dip rarely shows up in posts about AI. Mine lasted the whole month of March.
- March. Claude Code, Codex and the company's first SDR. Output drops while I learn.
- Spring. Skills, then loops.
- June 20. First commit of my system.
- July 11. An afternoon in a park, and the system takes shape.
- September. Testing harnesses, Jev, Muse and Instinct.
The estimate, point by point
| When | Estimated output |
|---|
The concepts that bent the curve
After tokens, the concepts came one after another. Skills: instructions an agent loads only when a task needs them, so it doesn't carry everything all the time. Loops, when they became the thing everyone was trying: an agent works, checks its own result and goes again until it passes. And context, the lesson from the GPT years, finally with a name.
Along the way came the rest. Connectors that let an agent read my email and my calendar. Subagents that split a big task into pieces that run in parallel. A plan written by one model and carried out by another. And version control, so every change has a date and can be undone.
From Canva to connectors
The tools tell the same story. At the start it was Canva and little else. Iridian added ads and dashboards, Boston added a CRM and R, and Digitech added Salesforce and Sales Navigator. Then 2026 turned tools into connectors: Apify and Apollo to find companies and people, Gmail, Calendar and Drive wired into the agents, and NetSuite and Python for the numbers. In May I learned Remotion and Higgsfield, for video.
| Design and content | Canva | CapCut | Remotion (May)Higgsfield (May) | ||||
|---|---|---|---|---|---|---|---|
| Ads and analytics | Meta AdsGoogle AdsLooker Studio | GA4 and Search Console | |||||
| CRM and sales | ActiveCampaign | HubSpot | SalesforceSales Navigator | ApolloNetSuite | |||
| Data and automation | R | ApifyPythonGmail, Calendar and Drive | |||||
| AI and agents | ChatGPT and GPTs | Claude CodeCodexJev | |||||
| Year | 2020 | 2021 | 2022 | 2023 | 2024 | 2025 | 2026 |
The learning curve turns
By June the work lived in plain files I could track. On June 20 I made the first commit of what became my system: a baseline to follow every project, task and deadline I had in one place.
Then came Saturday, July 11. I sat in a park from one to three in the afternoon and wrote the first prompts that really mattered, the ones that set how the whole system would work. I recorded a few videos that afternoon. Everything I run today started there.
Two days later I sat down with my father, Elver, and showed him how to write a prompt for a real project: full context, and a clear list of what was missing. He wrote his own and ran it in Claude Code, alone, on his own computer. That's when I knew the curve had turned. I could explain what I'd learned well enough for someone else to use it.
What stayed
The two early lessons ended up as rules. From the GPT years: what's current and short gets read every time, and what's long and historical gets looked up only when it's needed. From Iridian: nothing goes out with my name on it until I've read it.
Today the work sits in plain files that the agents and I both read, and every correction goes back into them, so a mistake fixed once stays fixed.
Where I am now
Right now I'm testing the different harnesses and working out what each one is best for. In mid-September TypeSafe AI released Jev, a model built for fast decisions inside agents, and I'm testing where it fits in my workflows. I'm also refining my own personal assistant, built on Claude Code with the context of my work, and comparing it with Meta's Muse and with Instinct.
The rest of my time goes to two things. One is understanding what AI is doing to the economy at a larger scale, so I can find use cases that really raise productivity. I'm a little obsessed with that search. The other is software engineering. This wave asks for technical knowledge I didn't have, so I study it every day and fill the gaps one at a time.
