artformoney/claude-close-skill

claude-close-skill provides a slash command called /close for Claude Code that ends a working session cleanly by handing its important state to whatever session comes next. The idea is to work in short, fresh sessions rather than keeping one long running chat alive for weeks, since a chat that has been running that long ends up dragging around old dead ends and pasted error output, which makes answers feel worse even though the underlying model has not changed. It also addresses the surprise cost of coming back to an old conversation after a break, since a long idle chat loses its cheap cache and a tiny message like a simple greeting ends up reprocessing the whole history. The skill's answer is to treat a chat window less like a database. Before writing anything to disk, the close routine has the model debrief with you in the conversation itself, describing what it is least sure about and what might be lost once the context window closes, so your response can shape what actually gets recorded. It then rewrites a single briefing file capped at 60 non blank lines, describing the project as it stands right now rather than appending to a growing pile of dated notes, appends a short entry to a running session log, and files anything more durable into the project's own permanent documents. It can also run project specific closing duties defined in a separate file, each one gated on whether that duty's condition actually happened during the session, and it notices when a previous session ended without closing properly by checking the gap at the start of the next one. Installing it means copying one global skill file into your Claude configuration folder, then wiring a session start hook and a pointer in each project either by accepting the skill's own offer to set it up or by copying an example settings file. A macOS specific date command is used for the staleness check, which needs a small edit to work on Linux. The project has no stated license in the shown README.