Skip to content
Sabin Mainali

My AI operating system - skills, hooks and memory

How I set up Claude Code as a working system for product, design, content and data work, with rules that are enforced automatically instead of remembered.

Outcome at a glance

Repeatable work runs through tested skills, and quality rules are enforced by automatic checks, so output is consistent from one session to the next.

Context

I do product, design, content, data, sales enablement and video work. Much of it now runs with Claude Code, Anthropic's AI coding and work assistant.

Problem

An AI assistant starts every session fresh. Out of the box it forgets decisions, drifts from house style, uses the wrong punctuation and burns time re-reading large files. Writing the same instructions again every day does not scale.

My role

I designed and built the whole setup and I use it every working day.

Approach

I moved the rules out of my head and into the system, in four layers.

  1. RequestI describe the task in plain language.
  2. Skills loadThe matching skill loads its tested method, for example the help center or screen design skill.
  3. Hooks guardAutomatic checks run on every action and block output that breaks a rule.
  4. VerifyNothing is reported as done until it is checked in the real product, data or browser.
How a request runs through the system

What I built

  • Skills. Packaged methods for UI generation, support articles, motion video and brand systems. Each one carries its own checklist.
  • Hooks. Automatic guards that run on every action: one blocks typographic punctuation, one blocks invisible Unicode characters, one stops unnecessary screenshots, and one flags topic drift.
  • Subagents. Small helpers that search or review in their own context and return only the conclusion.
  • Memory. A file-based memory of preferences and project facts that carries across sessions.
  • A knowledge graph. A graph over company documents, so answers come from a focused query instead of reading every file.
  • Token discipline. Rules that keep large files and screenshots out of the context unless they are the answer.

Outcome

The quality rules I care about are enforced by the system, not by my attention on a given day. New work starts from a tested method. The same setup now powers the help center pipeline, the design system checks and the video workflow.

What I learned

Most of the reliability comes from the rules, checks and memory around the model. A chatbot answers. An agent finishes.