session-lifecycle
Every session has a beginning, middle, and end. Design for session boundaries instead of being surprised by them.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
메뉴
Every session has a beginning, middle, and end. Design for session boundaries instead of being surprised by them.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Pause and reset. Gather what's scattered, commit what's ready, prepare for what's next.
Bootstrap a multi-agent project. Define the meta-process, resolve every open question, close the circle.
One agent flies high. Holds strategic context, names the unnamed, never descends into implementation.
Manage a multi-agent tmux workspace. Status, allocation, prompt drafting, cleanup.
Propose tmux lane allocations for pending work. Status view + assignment plan + sequencing logic.
Show the current state of all tmux lanes. Fast, visual, no ceremony.
| name | session-lifecycle |
| description | Every session has a beginning, middle, and end. Design for session boundaries instead of being surprised by them. |
Every session has a beginning, a middle, and an end. Treat each one deliberately, or entropy will treat them for you.
A protocol for the arc of a working session with an AI agent. Not a rigid checklist, but a set of practices that prevent the two most common failures: starting without context (cold start amnesia) and ending without capture (lost work).
The deeper point: context windows are not infinite, memory compaction is not optional, and sessions end whether you plan for it or not. Design around these constraints instead of being surprised by them.
This covers the full arc. For a lighter mid-session pause, see Gather.
Load persistent context. Read memory files, project state, any handoff notes from previous sessions. Don't trust what you remember from last time: trust what's written down.
Check for in-flight work. Did the last session leave something running? An eval, a build, a long process? Harvest those results before starting new work. Orphaned outputs are wasted compute.
Orient. Read the roadmap, the task list, the current priorities. Ask: "What matters most right now?" The answer determines the session, not momentum from last time.
Verify the environment. Is the system in the state you expect? Running processes, git status, server state. Surprises discovered mid-session cost more than surprises discovered at the start.
Log as you go. Record experiments, decisions, and results immediately, even failures. If it's worth trying, it's worth recording. The log is for future-you, who won't remember why you tried that thing.
Manage compaction deliberately. In long sessions, context will compress. Don't let auto-compaction surprise you. When it's approaching, run the handoff cycle yourself:
This turns compaction from a threat into a tool. Each cycle forces crystallization, and the handoff log becomes an artifact that future sessions can reference.
Circuit breaker. If multiple sessions have been spent on meta-work, process, or infrastructure without producing primary work output, force a reset. Process exists to serve the work, not the other way around.
Capture in-flight state. If anything is still running: record what it is, where its output goes, what to do when it finishes, and any non-standard system state.
Update persistent memory. Stable findings, new patterns, changed priorities. Future sessions start by reading this. Make it accurate.
Name what's unfinished. Explicitly. "Continue X" is better than silence. "Blocked on Y, try Z next" is better than "continue X." Unfinished is fine. Unnamed is not.
Clean the ground. Commit changes. Sync if appropriate. Leave the workspace in a state where the next session, whether it's you or someone else, can start clean.