Conversation Summary

This document summarizes a conversation about working with an LLM.

The conversation started as a question about interaction patterns.

It ended with changes to this skeleton.

The Pattern

The unit of work is a work item, not a conversation.

A prompt is a work order, not a whitepaper.

Group strongly related items into an episode with one goal and one acceptance.

Phrase risk items as questions and put them first.

Treat the TODO list as a queue of episodes.

Decouple human attention from the model’s process.

Dispatch work asynchronously and review artifacts later.

Use git as the message channel.

Keep the human as the synthesis point and final authority.

Use worktrees to isolate LLM work.

Separate intent from record.

Use notebooks for exploration, not for work items.

The full pattern is in 00-pattern-of-interaction.md.

The Skeleton Changes

Preferences

Static-Site Discussion

The skeleton demonstrates static-site generation without an SSG.

scripts/site-build.sh already wraps site.in/*.txt in literal HTML pages.

A minimal Markdown-to-HTML transformation is tempting.

It cuts close to replicating an SSG.

A hand-rolled transformer owns escaping, code fences, and Mermaid blocks.

Mermaid can render client-side in the browser, keeping the host rsync-only.

The client-side Mermaid rendering choice is accepted.

A Markdown-to-HTML transformation still needs a parser, which is a build-time dependency.

Near-literal output serves developer readers.

Non-developer readers want styling and navigation, which is SSG territory.

The decision is open and tracked in TODO.md.

Open Questions

The open questions live in TODO.md.

They include Jupyter, static-site generation, the DELTA protocol, and the tool universe.