How I work with AI
I use AI every day, and I am careful about what I let it decide. It makes me faster at design, documentation and code. It does not replace understanding the project.
Three surfaces
Think
Run discovery, analyse requirements, compare architectures, write documents.
Claude Chat and Claude Cowork
Build
Write and test code, deploy, manage the repository.
Claude Code
Reach
Read configuration in systems where no API access is granted.
Claude in the browser, with me present
How context survives
AI tools forget between sessions. I keep the durable record in the repository: a project instruction file that states the standards and constraints, and a decision log that records what was decided and why. Any session, human or AI, starts from the same truth.
What I have built for it
- A packaged Claude Code plugin with four project skills and two hooks: one checks code after every edit, one blocks a commit that contains secrets or failing tests.
- Session starters that give a new AI session the right context for one workstream.
- A plan for which connectors to wire in, and which systems have none.
- Requirements in Notion and stories in Linear, both kept in step with the code by requirement ID.
Rules I keep
- AI drafts, I decide
- Nothing customer-facing, financial or legal goes out without my review.
- AI never handles credentials, money or identity
- Those stay with a person and with the platform's own controls.
- Generated code meets the same bar as mine
- Static analysis, tests, review.
- You need to understand the whole journey to use AI well
- From requirement to deployment. I do.
Anthropic certifications
- Claude 101
- Claude Platform 101
- Claude Code 101
- Claude Code in Action
- Claude Cowork