Each model does the job it is best at.
A coding agent handles code, a fast model handles research and review, and the strongest model is kept for planning and judgement.
Independent workflow · set up October 2026One lead model.
One lead model.
Four specialists.
How I use AI to build and maintain this portfolio. One model plans and checks, specialist helpers each do one job, and I decide what gets published.
The team
Who does what.
Me
Anthropic
The lead
Claude Opus 5.5Plans the work, hands out jobs, checks every result and reports back to me.
hands jobs to
- OpenAI
Coder
OpenAI CodexWrites the code.
- New pages and features
- Changes across many files
- Hard bugs
- Anthropic
Researcher
Claude Sonnet 5.5Looks things up.
- Companies and job market
- Dissertation reading
- Finding code in the repo
- Anthropic
Reviewer
Claude Sonnet 5.5Checks for mistakes.
- Reads every change
- Runs lint, type and build checks
- Flags broken content
- Anthropic
Writer
Claude Sonnet 5.5Writes the words.
- Site and project copy
- CV bullets and cover letters
- House style, no hyphens
Example
How one job moves through the team.
Adding a new case study page, from request to live site.
- Opus
Plans the job
Works out what a new case study page needs and which files it touches.
- Researcher
Gathers context
Pulls the project notes and shows how existing case pages are built.
- Codex
Builds it
Receives a full written brief and writes the page code.
- Reviewer
Checks it
Runs the site checks and lists anything that is broken or unclear.
- Opus and me
Signs it off
Opus fixes what was flagged and reports back. I read it before it goes live.
The builder never marks its own work.
Every change Codex makes is checked by a separate reviewer before it is committed, the way a second analyst would check a model.
Research and review run at the same time.
Helpers that do not depend on each other work side by side, so a task takes the time of its slowest step rather than the sum of them all.
I stay responsible for what is published.
The rules the team follows are written down in the repository, and nothing reaches the live site without passing the build checks and my review.
What this does not claim
The AI does not choose my work or my conclusions.
The models write code and drafts when asked. I choose what to build, read changes before they are published, and the ownership of each case study is stated on its own page. When this page launched in October 2026, Codex was not yet connected, so Opus built it directly with my approval.