Recovering from a runaway agent: rollback patterns
Sometimes an agent goes way past scope and modifies things it shouldn't have. Here's the rollback playbook — git-based, fast, low-stress.
11 posts
Sometimes an agent goes way past scope and modifies things it shouldn't have. Here's the rollback playbook — git-based, fast, low-stress.
Claude Code's memory feature is powerful but easy to misuse. The pattern that scales — what to put in global memory, what to put per-project, what to never persist.
Five parallel agents produce a lot of output. The trick is summarizing what mattered without reading every commit. Here's the daily routine.
cmux turns one terminal window into a tile of independent agent sessions. Here's how to wire it up, name your sessions, and integrate it with a file manager so context never gets lost.
Naming AI sessions feels trivial until you have 50 of them. The convention that scales — and the patterns that break.
Every new AI project needs the same handful of bootstrap files. Skipping them costs an hour each session. Here are the seven templates that pay back immediately.
Inheriting code is easier with an LLM if you ask the right questions in the right order. The 4-step process that surfaces the architecture in under an hour.
When you're using an LLM to read papers, summarize sources, and write a synthesis, four file-manager panes beat 14 browser tabs. Here's how to wire it up.
CLAUDE.md is the most-undervalued AI productivity tool. A good one saves hours per week; a generic one is dead weight. Here's what makes the difference.
Stop pasting the same instructions into every chat. Here's a battle-tested layout for a personal prompt library that you can grep, version-control, and share between Claude Code, Cursor, and Codex.
A directory layout, naming convention, and pane-routing strategy for running multiple Claude Code sessions on macOS without losing your place.