Skip to content

AI Lab

This page explains the AI setup I use to do my work, in plain language. It is built on Claude Code, and it turns a general AI assistant into a system that follows my rules and checks its own output.

"A chatbot answers. An agent finishes."

How a task runs

Every task passes through the same five stages. The guard and verify stages are what make the output reliable.

  1. Request

    I describe the task in plain words.

  2. Skill

    The matching method loads with its checklist.

  3. Guard

    Hooks check every action as it happens.

  4. Verify

    The result is checked in the real product, data or browser.

  5. Done

    Only verified work is reported as finished.

From request to verified result.

The five parts of the system

Each part solves one weakness of a plain AI assistant.

Skills

Like a recipe card

A skill is a written, tested method for one kind of job. When a task matches, the skill loads and the work follows its steps and checklist.

  • Help center articles: voice rules, click-proof steps, personal data masking
  • App screen design: spec, generate, brand lint, fix
  • Product video: storyboard first, then a visual QA loop
  • Published pages: search and AI-search structure applied by default

Hooks

Like a spell checker that cannot be switched off

A hook is an automatic check that runs every time the assistant acts. It can block the action and explain why, so a rule never depends on memory.

  • Punctuation guard: blocks typographic characters such as curly quotes
  • Hidden character guard: blocks invisible Unicode in files
  • Screenshot guard: stops screenshots unless pixels are the answer
  • Topic drift guard: flags work that wanders from the request

Subagents

Like sending an assistant to the archive

A subagent is a helper with its own workspace. It searches or reviews a large amount of material and brings back only the conclusion.

  • An investigator that finds where something is defined
  • A reviewer that checks a change for defects
  • A builder that makes a small, well-scoped edit

Memory

Like a notebook that carries between days

Memory is a set of small files with preferences, decisions and project facts. Each new session starts with what matters instead of from zero.

  • How I want work reported back
  • Standing project decisions and constraints
  • Pointers to the right dashboards and documents

Knowledge graph

Like an index at the back of a book

A knowledge graph maps company documents and how they connect. A question pulls a small, relevant slice instead of reading every file.

  • Fast answers across product, support and sales documents
  • Relationships between features, articles and decisions

What this changes in practice

Questions clients ask

Short answers about how I work with freelance clients.

What is Claude Code?

Claude Code is an AI assistant made by Anthropic that works directly with files, tools and the command line. Sabin uses it for product, design, content, data and video work, not only for writing software.

What is the difference between a chatbot and an AI agent?

A chatbot answers a question and stops. An agent finishes a task: it plans the steps, uses tools, checks its own work and reports when the job is done. Sabin builds setups that behave like agents.

Can Sabin set up a system like this for my team?

Yes. The AI workflow automation service starts with an audit of one real workflow, then adds the skills, prompts and automatic checks that fit it, and ends with a short handover guide for the team.