Jorge Castro · Applied AI for software development
Work with AI agents while keeping control of your code.
I write about what I actually use when I work with coding agents: the rules I set before letting one touch anything, how I manage context once a project grows, and where I got it wrong, which tends to be the useful part. I decide and run this site together with ARK, my decision system, and the final judgment stays mine.
Delivered in Spanish by email. Free. Unsubscribe in one click.
Latest articles
Start here
The instructions I start with, how I manage context, and a case that is still running with my team.
Qué hay realmente dentro de mi CLAUDE.md
Las reglas que le pongo a un agente de código antes de dejarlo tocar nada, y el fallo concreto que previene cada una. Incluida la que más trabajo me ha ahorrado.
Más contexto no es mejor contexto
Le di a un agente acceso a todo el repositorio para arreglar tres líneas. Falló, y no por falta de información. Las cuatro reglas que escribí después para administrar el contexto como el recurso escaso que es.
De horas a minutos: Cómo automaticé el cierre de Sprints con IA y n8n
Reescribiendo el rol de quien lidera el equipo: cómo automatizar los reportes nos devolvió tiempo y le dio autonomía al equipo.
About
Fifteen years building software, and still learning how to decide.
I'm Jorge Castro. I started supporting systems that were already in production, then founded a company I ran for nine years, and today I lead a technical team. I use AI to build and run real software, and I write down what worked, what did not and why, errors included.
Case study
A sprint wrap-up with AI and human review
Every sprint close used to take my whole Friday copying numbers from one tab to another, so I wired the team tools to n8n and a language model to assemble the report, which I still review before it reaches anyone. It is still running, and it does a good deal more than the version in the article.
Read about the process and its limits- Sprint data: Jira tickets, CI/CD logs, production incidents and pull requests.
- n8n gathers it and a language model summarizes progress and deviations.
- I review the report before it reaches stakeholders.
- The team checks the dashboard on its own and fixes tickets before the review.
How this site runs
A site co-managed with AI
ARK is the Markdown decision system I consult before deciding what to publish, what to measure and what to drop, and its main job is to argue with me when what I propose collides with something I already decided. It takes part, and it never publishes or decides on its own.
How ARK works- Every article declares whether ARK was consulted or assisted. Without that label, authorship is human only.
- Decisions that are hard to reverse, such as the site's public identity or the scope of its redesign, are recorded with context and alternatives before they are carried out.
- ARK checks what I propose against what I already decided. It never starts an action: the last word is mine.
Resources
A free file you can start with today, and a paid system for when context and decisions begin to get lost between sessions.
Personal context system
ARK
A system for preserving context, challenging decisions and keeping agent work under human judgment.
Explore ARK
Working template
Base CLAUDE.md
The rules I use to start a project, annotated from experience and ready to adapt.
Explore CLAUDE.md
Let's discuss a concrete idea.
If something here connects with a problem you are solving right now, write to me and tell me the context, which is the difference between a generic answer and a useful one.
Alcance
Haz exactamente lo pedido. No amplíes el alcance sin una razón declarada.
Implementa el cambio correcto más pequeño.
Por qué está: es la regla que más trabajo me ha ahorrado.
Límites
Nunca modifiques un test para que pase. Si un test falla, el sospechoso es el código.
The base CLAUDE.md
The instructions I use to begin projects with agents, annotated rule by rule and ready to adapt.