A self-portrait with AI: what my data says about how I learn, write and code

For months now I've worked every day with an AI agent in my terminal. I write specs, code, courses and even this blog with it. At some point I asked myself the uncomfortable question: how much of all this is actually mine? I didn't want to answer with an opinion. I wanted to answer with data.

So I built a pipeline that reads my Claude Code history and my two GitHub accounts, and I set it against what the world is saying about AI, writing and authorship. This post is three things at once: an honest self-portrait, a sample of how I work (it's what I teach in MAPA), and a method you can copy at the end.

One disclaimer up front: this post is co-authored too. Claude wrote the draft from the data, and I reviewed it, corrected it and decided what gets published. I approved every quote from my prompts one by one. If that feels like cheating, stick around for the tensions section.

The context

Before looking at myself, I looked outward. The debate is full of big numbers, and almost all of them are self-reported. Sundar Pichai said 75% of Google's new code is generated by AI and reviewed by humans (source). People keep repeating "90%" for Anthropic, but an outside analysis by Redwood Research puts it at ~50% of merged lines (source). These are executives' numbers, each measured its own way, and they can't be compared with each other.

There are more modest outside measurements. SemiAnalysis estimated that ~4% of daily public GitHub commits carry Claude Code's signature (source). In Stack Overflow's 2025 survey, 84% use or plan to use AI, but 46% distrust its accuracy (source).

On productivity and quality, the evidence is awkward. In a controlled experiment by METR, experienced open-source developers took 19% longer with AI while believing they were 20% faster (source). GitClear measured 8x more duplicated code blocks across 211 million lines (source). And an experiment by Anthropic itself found 50% comprehension with AI versus 67% without it (source).

On writing, the ICMJE is clear: AI can't be listed as an author because it can't be accountable for the work (source). And a Cornell study showed that AI suggestions pull writing toward Western styles: Indian writers accepted 25% of suggestions versus 19% for Americans (source).

Around those numbers there are camps: pragmatic enthusiasts, evidence skeptics, people defending agentic engineering with real review, maintainers shutting the door, worry about juniors, and the old question of authorship. What I couldn't find was verifiable data from Latin America. That's why this exercise, even if it's just one person, seems worth doing.

The map highlights the debate's concepts that also show up in my prompts. There are only four out of 23: agentic engineering, Co-Authored-By, AI disclosure and code review. I talk to AI a lot, but almost never about AI.

The timeline

The calendar runs from November 18, 2025 to today: 81 active days and 2,600 prompts across 210 sessions. Usage is anything but even. May had 20 active days and 795 prompts; September, just 3 days and 10 prompts. My two busiest days were October 14 and 15 (116 and 167 prompts), writing exactly this, and the third was August 19, with 113.

What I wanted to know was whether conversations turn into published code. Sometimes they do, sometimes they don't. In March I wrote 244 prompts and my new account shows no commits that month. It was a month for exploring, not shipping.

Where I put AI

I grouped prompts by domain. Web products take 1,133, more than four in ten: this site (sergiomonsalve.com), bichongos.store, dharmastudio.art, InteFin and sama-mobile, among others. Then come experiments (569), agents (335, with brayan and HER), mushrooms and IoT (330, with Bichongos and mobile-esp32) and education (233, with medmap, New-Old-School and stemedu).

Private repos count toward the totals but never appear by name. That's why "experiments" weighs so much while showing a single public repo.

How I work with it

This comes from the full sessions Claude Code keeps, a sample from March to October, not the whole history. By far the tool I use most is Bash: 3,642 calls out of 5,801, almost two in three. Research comes next: 500 web searches and 446 page fetches. The browser (Playwright and Chrome) adds 260 uses, and orchestration (subagents, skills, artifacts) another 467 calls.

One detail I like: the AI asked me 71 clarifying questions before moving on. It isn't all blind delegation.

As for models, Opus 5.5 gave 8,739 of the 11,755 responses in the sample, followed by Opus 5 with 2,469. My reading: I delegate execution and keep the judgment. The agent drives the terminal; what to build and when to stop is my call.

Two eras

I have two GitHub accounts and, without planning it, each one is an era. The old one (smonsalve) runs from 2013 to 2025, with a few commits from 2026. The new one (serandmoncas) started in February 2026.

On the old account I have 417 commits of my own across 73 repos, over thirteen years. On the new one, 1,460 commits across 31 repos, in about eight months. The change in writing is even starker: my median commit message used to be 13 characters. Today it's 137, or 87 without the co-authorship trailer. I used to write "fix". Now there's a paragraph explaining what changed and why, even if the AI drafts much of it.

The co-authors changed too. In the old era five people signed with me, classmates and a business partner, adding up to 101 commits in my public repos. In the new one there are no people at all: there are six Claude models taking turns. Sonnet 4.6 dominated from February to June, Sonnet 5 took July, Opus 5 arrived in August, and Opus 5.5 signs all 286 co-signed commits in October. Opus 4.7 and Haiku 4.5 show up occasionally.

Of the new account's 1,460 commits, 903 carry Claude's signature: 62%. Month by month it hovers around half from May to July, rises to 60% in August and hits 97% in October (286 of 296). Commit types include things that aren't software: cultivation, training, content, firmware. AI is already in everything I do, not just the code.

And one detail I didn't expect: Claude isn't only a co-author. In an education repo (New-Old-School) it's the main author of 69 commits.

Three caveats, so I don't fool myself or you. First, the trailer measures who made the commit, not who came up with the change. It's a floor, just like SemiAnalysis's ~4%. Second, the number depends on the workflow: in October Claude commits by default, so the jump to 97% says as much about my setup as about how I work. Third, only public repos are visible on the old account: the work I did for companies in those years doesn't show up, so the baseline underestimates how much I coded before AI.

The ecosystem

Claude Code is the only tool I can measure. The rest are self-reported: Claude.ai for long conversations and artifacts, ChatGPT for quick questions and the API, Ollama for local agents, and Gemini and Cursor now and then. The matrix labels this clearly, because a measurement and a perception are not the same thing.

What I master vs. what I don't

I wanted to know which concepts I actually master and which ones I just name. I took the 200 terms in the Experiencias IA glossary and crossed them on two axes. Spoken: the term appears in my prompts as an instruction. Built: it appears in my repos' code or in my stack backed by code. If a term only shows up in questions ("what is", "explain"), I mark it as asked.

The classifier is lexical: it measures use, not understanding. That's why I reviewed the table by hand and corrected 3 of the 200 concepts.

The result is humbler than I expected. Only 4 concepts land in both: Superpowers, Claude Code, agents and automated testing. There are 17 I only speak, like handoff, MCP, token, CLAUDE.md or skills. Only one is built without being spoken: embeddings. And 178 appear on neither axis, including prompt engineering, context engineering and RAG. RAG is on my declared stack, yet it doesn't appear even once in my prompts.

Six concepts show up in questions, including harness engineering (3 times) and handoff (2). ADR is the only one I've only asked about, never used as an instruction. To me, a concept I only ask about is an honest signal: I'm learning it, I'm not using it yet.

The path backwards

Then I asked myself what I had to learn to reach the stack I use today. I rebuilt a knowledge graph with 52 nodes and 74 dependencies: 35 with evidence on GitHub, 7 self-reported and 10 inferred foundations I approved by hand. Each date comes from the first commit that backs it.

Rewinding it gave me two readings I didn't expect.

First: my current work with agents doesn't depend only on programming. Evaluating agents builds on experimental design (oTree, 2017) and statistics, and the MAPA method comes from teaching programming (2016) and producing educational content (2014). What looks like "knowing AI" today is, to a large extent, knowing how to teach and how to design experiments.

Second: one of the things I delegate most is documentation (289 docs commits on the new account), and that's a skill I already had. My first repo with technical writing in LaTeX dates from 2014. I'm not handing the AI something I never learned. I'm handing it something I know how to review. I keep thinking about Anthropic's 50% versus 67% comprehension result, and I think the difference matters: delegating what you've already mastered is not the same as delegating what you're still learning.

How I talk to AI

Leaving out commands (the ones starting with /), there are 2,475 written prompts. Their median length is 18 characters. Only 13% are questions, and 80% are in Spanish. That's the clearest finding of the whole exercise: I don't converse with AI, I give it short orders. "Commit it." "Continue." "Update the backlog." Long instructions exist, but they're the exception, and they almost always show up when I start something new.

Here are some of those exceptions, with only the month:

act as a software architect and help me improve the functionality and UX of this project

— December 2025 (translated from Spanish)

help me create a backlog with the tasks mentioned so far with description, priority, prerequisites, and whenever we mention a new task or a dependency of some task create it and define it in the backlog

— June 2026 (translated from Spanish)

commit it, remember to add a good definition for each term, maybe an example, references, links, youtube videos, books, or important references, maybe it should lead to another subpage with the term and all those fields?

— July 2026 (translated from Spanish)

update the backlog and continue with the next priority

— July 2026 (translated from Spanish)

help me do harness engineering on this project create a backlog with the tickets done so far and if new tickets come up propose them

— August 2026 (translated from Spanish)

help me apply the curioso methodology to this topic and to the presentation help me using the [Claude visualization] tools we're currently developing and learning to use and whatever you think is worth it from other projects

— October 2026 (translated from Spanish)

what if I wanted the viewer to build me a timeline of what happened in the evolution of a topic: for example I want to learn about aviation give me an index showing me the big milestones in aviation

— October 2026 (translated from Spanish)

There's a pattern: the backlog comes up again and again. I ask the AI to keep track of the tasks, and I decide the order. That's what delegating execution without letting go of judgment looks like.

Tensions

I don't have settled answers. I have five questions that cross the outside world with my own data.

Co-author or assistant? Git gives me Co-Authored-By, but Assisted-by also exists, and it names assistance rather than co-authorship. The ICMJE says AI can't be an author because it can't be accountable for the work. Yet 903 of my 1,460 recent commits carry its name as co-author. Am I using the right word?

The trailer is a floor, not a measurement. SemiAnalysis's ~4% and my 97% in October measure the same thing: who left the signature. In October Claude leaves it because that's how I configured it, not because it did more than in May. If I turn the trailer off, the number collapses and the work stays the same.

The METR paradox, applied to me. I feel much faster. I went from 417 commits in thirteen years to 1,460 in eight months, though that count includes my commits plus the ones Claude co-signed or wrote: it measures the volume of published work, not what I typed. But METR found people who felt 20% faster and were 19% slower. My data measures volume, not speed or quality, and I have no control group.

Review as a new form of authorship. Simon Willison summed it up in an uncomfortable line: "But I'm not reviewing that code". If my prompts have a median of 18 characters, my contribution isn't in what I write but in what I review, accept or reject. And if I don't review, GitClear has already measured what happens: 8x more duplicated code.

My own voice. 80% of my prompts are in Spanish, and I live in the mountains of Antioquia, Colombia. Cornell showed that AI suggestions push toward Western styles. Claude wrote this post from my data, and I corrected it; this English version is itself an adaptation of the Spanish one. Does it sound like me? You tell me. That's why I declare it as co-authored.

Do it yourself

The whole analysis lives in the site's public repo, under scripts/ia-autorretrato. To run it on your own data:

The only thing that gets published is data/ia-autorretrato.json, with aggregates. Before writing it, a privacy filter blocks the output if it detects text from your prompts, private repo names, people's names, emails or local paths. The same check runs on the text of this post before it's published.

If you want to learn to work like this, with agents, with method and without losing your judgment, that's what I do in my courses. And if you'd rather start with the concepts, the glossary is in Experiencias IA.