Autre~80 · IA en attente

Two Claude Code skills for long sessions: a real handoff prompt to a fresh session, and a "where are we / what do you need from me" panorama

r/ClaudeCodeu/Expert-Paramedic93831 octobre 2026

Capture du projet

Analyse IA en cours de préparation : les informations ci-dessous proviennent de la détection automatique.

Résumé

This post was entirely written by Claude execpt for this very line, just sharing some stuff that might help a lot of people i've seen in here, but i'm too lazy do this on my own, kek. Enjoy, peace! If you run long Claude Code sessions you probably know both of these: • Context fills up. Auto-compaction loses detail, s…

Afficher le post original
This post was entirely written by Claude execpt for this very line, just sharing some stuff that might help a lot of people i've seen in here, but i'm too lazy do this on my own, kek. Enjoy, peace! If you run long Claude Code sessions you probably know both of these: • Context fills up. Auto-compaction loses detail, so you open a fresh session, type "continue where we left off", and the new session has no idea what "where" is. It re-reads half the repo and asks you things you already decided an hour ago. • You get lost. The session goes deep into something technical, Claude says "should I go with A or B?", and you honestly don't know what either option means for your project. I wrote two small skills for these and have been using them daily for a while. Sharing in case they help someone: /session-handoff writes a self-contained, copy-paste prompt (in a code block) for a brand-new session. What makes it work better than "summarize the chat": • it asks you the pending decisions before writing, so the new session doesn't stall on turn 1; • it re-checks the real state (git status, last commit, running jobs, test results) and re-opens every file:line it cites instead of going from memory; • it marks what's verified vs inferred; • it carries your working style over (subagents, TDD, plan-first, etc.); • fixed sections: where things stand, settled decisions, scope, ordered open items with "done when", gotchas, and the concrete first action. /panorama is a read-only "big picture" in three layers: what this session did and what's blocked, how that fits the project, and each decision that's yours, with what actually changes if you pick each option, a recommendation, and whether it's reversible. It ends with a one-line answer template ("1: A / 2: yes") and then stops and waits. It also checks whether docs/handoff notes are stale against git history, and says "I don't know" instead of inventing. Repo (MIT, just two SKILL.md files): https://github.com/gildeshiro/claude-code-skills Feedback welcome, especially if the handoff misses something your next session always ends up asking for.