Act on this, mostly signal

Six Claude Code practices from about 20 unattended home jobs

What to do

Shorten instructions, start fresh sessions per step, and tier models for unattended jobs.

Why

The non-developer author reports these practices, and Claude sometimes declares work done too early.

What the research found

A non-developer shares six Claude Code practices from running about 20 timed jobs on a home machine: shorter instructions, fresh sessions per step, script-enforced rules, pruning homemade skills, tiered models, and locked-down unattended jobs, plus an open problem of Claude declaring done without testing.

Claims and how well the evidence supports them

The author runs Claude Code on an old Dell OptiPlex with Lubuntu and a Max plan with about 20 timed jobs 98%
Of 12 skills and 7 hooks, only one had been replaced by a built-in feature 99%
A hook blocks saying done when code changed and nothing ran afterward 98%
Haiku finds, Sonnet writes from a brief, and Opus builds 99%

Evidence

reddit.com https://reddit.com/r/ClaudeAI/comments/1xcob2x/not_a_developer_running_my_life_admin_and_a_few

Actionability rated medium.