ARK
The decision system I consult before deciding
ARK is a Markdown knowledge system that a coding agent reads before advising me. It keeps the context I forget, checks what I propose against what I already decided, and leaves the reason in writing. The final judgment stays mine, and that boundary is written inside the system itself.
What you buy is a Markdown system that integrates with your coding agent, in Claude Code or Codex, plus an onboarding to fill it with your own context. It is not a copy of my repository.
What it is
- Plain text files. No database, no dashboard, no automation.
- A body of principles, a roadmap and the active campaigns, loaded by the agent before it answers.
- Councils that apply public analytical frameworks to a concrete decision, without inventing quotes or attributing opinions.
- A dated record of the decisions that are expensive to reverse, with their context and the alternatives that were dropped.
What it is not
- It is not an autonomous agent. It never appears as the initiator of an action.
- It is not a service or an integration: it connects to nothing and runs on its own schedule for nothing.
- It does not replace judgment. It forces judgment to be written down, which is less comfortable.
- It never marks anything as verified on its own initiative. A person confirms that.
How it runs
Four steps, always in the same order
- Classify the request and load only the relevant context, instead of dragging in the whole repository.
- Separate facts, inferences, assumptions and pending evidence before recommending anything.
- Surface the contradiction when what I ask collides with the principles, the priorities or the family constraints.
- Record the decision when it is hard to reverse, so that the version of me six months from now knows what was ruled out.
An example
A decision receipt, in a fictional case
Caso enteramente ficticio. Cifras, personas, fuentes del caso y desenlaces son inventados para evaluar el método.
Lucía dirige Bruma, una empresa de dos personas que vende una herramienta de inventario a talleres mecánicos.
Le llegaron dos caminos posibles y solo alcanza para uno.
El recibo propuesto
- Decisión propuesta
- Conversar con los tres talleres
- Resultado esperado, escrito antes de actuar
- Al menos dos de tres describen un costo concreto, en dinero o en horas
- Criterio de abandono
- Si ninguno nombra un costo, la hipótesis se reformula antes de construir
- Qué refutaría la premisa
- Tres conversaciones sin costo identificable
- Fecha de revisión
- Siete días, o antes si las tres se cierran primero
La elección fue correcta y el resultado quedó corto.
The proof
What changed in real work
- Sprint tracking and release management dashboards built with n8n and the Claude API, which removed six hours of manual work a week.
- Every article on this blog declares whether ARK took part and how. The absence of that label means human authorship with no assistance, and the absence reads at a glance.
- The decisions that show up in the articles came out of this system, including closing a product that failed its own success criterion.
What gets published and what does not
The line is written before publishing, not during
The repository holds real personal and financial information, so its contents are not published. What is public: the method, the structure by category, and results that can be verified from outside. What is not: finances, active campaigns, review records and the text of the decisions. If any of that appeared here, the system would have stopped doing the job it exists for.
One-time payment of 19 dollars. Personal licence. If it does not work for you, you get your money back, no questions asked.