Comment-concision mode. Comments must earn their place (a non-obvious why, or meaning a name can't hold) and stay to one line, so most functions carry none. Investigation details go to the commit message. Docstrings are exempt, and it never changes what the code does. Modes: off, new, tidy (default), plus /fp-minify:strip for a deliberate cleanup pass.
1 skill · 2 commands · hooks
Understanding-first mode. On exploratory tasks (refactors, design decisions, fuzzy scope) the model reads the code, then settles the real forks with you through the option-dialog before it writes anything, so you can debug and own the result. Trivial and fully specified work is built straight away. Modes: off, on (default).
1 skill · 1 command · hooks
Tags each load-bearing claim with its basis (🟢 verified, 🟡 inferred, 🟠 recalled, 🔴 guess) and ends the response with a one-line tally. A backstop hook checks verified tags against the tools actually used. Pure annotation, it never changes what the model does. Modes: off, footer, full (default).
1 skill · 1 command · hooks
Readable-prose mode ("sic 'em-dash"). In the prose the model writes (READMEs, docs, markdown, PR and commit bodies) it replaces the em-dash and its cousins with connectors every English reader can follow at speed, non-native readers included. Clearer, never dumbed down. Chat, code and comments are left alone. Modes: off, on (default).
1 skill · 1 command · hooks
Terminal-recall mode. When Claude Code stops to ask you something (a permission prompt, a question dialog, an agent blocked on input) it brings the terminal window to the front, unless it already has focus. Works on KDE, Hyprland, sway, X11, macOS and Windows, with a bell and desktop notification where no window can be raised. Pure mechanism. Modes: off, hint, on (default).
1 skill · 1 command · hooks