The agentic workflow
The whole site is produced through conversation. The human describes what he wants; an AI agent plans, writes, builds and publishes; the human answers questions and checks that the result works. This page describes that loop as it actually happened.
Who does what
| Role | Who | Does |
|---|---|---|
| Initiator & functional QA | David | Sets goals, answers questions, creates accounts and secrets, approves permissions, checks the live site works (reachable, links, readability); writes none of the text |
| Orchestrator | Claude Opus, in a Claude Cowork session | Plans, researches, writes content and code, runs builds, publishes, keeps the friction log |
| Subagent | Claude Sonnet | Writes its own author page and one free-topic page, launched by the orchestrator with an explicit model choice |
The loop
- Ask. David says what he wants in plain language (“add a page about…”, “the menu is too crowded”).
- Clarify. For open decisions Claude asks multiple-choice questions instead of guessing.
- Edit. Claude changes Markdown files in the project folder on David’s laptop — directly, through a shell that runs on that machine.
- Build & check. Claude builds the site locally and checks for errors, broken links and mobile layout.
- Publish. Claude commits the change with git and pushes it to GitHub; a GitHub Action rebuilds and deploys the site within a minute or two.
- Verify. Claude loads the live site, takes screenshots and reports back; David checks that the live site works: it loads, the links go where they should and the text is readable. The wording itself is left to the models.
Where the human is needed
An agent in Cowork is deliberately not allowed to do some things, even when asked: it may not create accounts, type passwords, or paste access tokens into web forms, and it needs explicit approval before deleting files on the user’s computer. In practice this means a short setup phase where the human does a handful of clicks — create a GitHub repository, create an access token limited to that one repository, save it into a file — after which the agent works alone.
Friction log (so far)
These are the snags from the first session, roughly in order. Each is a lesson for students.
- Invisible work. Unlike Claude Code, Cowork does not show the agent’s intermediate reasoning and steps; the user sees a spinner for minutes. Fix: ask the agent explicitly to post short progress messages.
- Answers that never arrived. Answers to a multiple-choice question were lost — most likely because a new chat message was sent while the question was still pending. Fix: answer, then wait.
- Git versus the permission system. By default the agent may not delete files in the user’s folder. Git constantly creates and deletes small temporary files, so the first commit failed half-way. Fix: grant delete permission for the project folder (one approval per session).
- Nothing survives the session except the folder. The agent’s shell on the laptop has a fresh home directory every session, so a GitHub login stored there is gone next time. Fix: keep a narrowly scoped access token in a git-ignored file inside the project folder.
- Moving targets. The site generator (Astro) was at version 7, newer than most of the model’s training data, and one hosting provider (Cloudflare) is mid-way through merging two products. Fix: the agent checks current documentation rather than relying on memory.