Advanced uses

Muse can also be the one holding the tools. Infrastructure, design, and data accumulation — the sit-down, long-horizon kind of work.

Operating the cloud from the CLI

Install Cloudflare's official CLIs (cf and Wrangler) into Muse's environment, and dashboard hand-work becomes commands. The key is permission design: API tokens are scoped to one account and only the permissions needed, in a form that can be revoked at any time. Separating production and laboratory accounts is a prerequisite.

Co-writing design documents

Big systems don't start as code; they start as a chaptered design document. For a system that tracks news as "stories", we wrote a full model document — concepts, architecture, how data is held, how it will be evaluated — then revised it under another AI's critique. When a document outgrows its patches, our house rule is to rewrite it whole rather than keep appending.

Designing how data accumulates

Decide how information is stored before it piles up. Each system owns its data; connections between systems go through declared types and a catalog. Judgments are never rewritten after the fact — they're recorded as they were when the outcome was still unknown. Done right, a system invented later can pull exactly the parts it needs out of years of accumulation.

A bench-and-production way of building

Compiling and verifying source code happens first in Muse's Linux environment, where trial and error is cheap. Only the proven procedure gets reproduced on my own PC (under WSL) as the "production" setup. Break things where breaking is cheap, first.

Designing background machinery

Combine scheduled runs (nightly, hourly) with condition watches (wake only on change) to build automation that runs silently. The important part is staying quiet when nothing is happening: notifications only for changes that matter.