Menu
Check memory for in-progress tasks and active TODOs at the start of each session. Use at the beginning of every new conversation.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Rules and style preferences for the usage of git. Use before you interact with git in any way.
Write a handoff document so a fresh session (or another agent) can resume multi-session work with full context. Use when the current session is ending mid-effort, context is running high, or the user asks to "hand off", "hand over", or prepare the next session/slice.
Checklist for wrapping up a session - persist memory, check TODOs, and ensure nothing is lost. Use before marking a task complete or ending a conversation.
Evaluate and refine a plan before implementation begins. Use once you think you are ready to present a plan or design to the user.
Analyse recent session transcripts to extract learnings, discover pain points, and persist them into memory/skills/rules. Run manually when the retrospective counter is due (session-end reminds you at 5+).
Reduce the amount of code in a changeset or file - remove duplication, dead code, redundancy, and over-verbose constructs - while preserving behaviour. Use when asked to tidy, shrink, deduplicate, DRY up, or simplify code.
| name | session-start |
| description | Check memory for in-progress tasks and active TODOs at the start of each session. Use at the beginning of every new conversation. |
At the start of every session, before responding to the user's first message, check memory for in-progress work using compact mode to minimise token usage:
read_graph(project="<repo-name>") to surface recent entities for the current workspace.compact=true:
search_all_projects(query="task", status="in-progress", compact=true)search_all_projects(query="task", status="planned", compact=true)
These are the only authoritative source for the task summary. read_graph only returns 10 recent entities from one project and will miss tasks in other scopes. compact=true omits observations to save tokens - you only need entity names and statuses for the summary.llm-prompts update --check to see if any tool sources have upstream updates available. If updates are found, mention them to the user.read_graph, search_all_projects, search_nodes, etc.) directly - they are ordinary tools in your catalog. Do NOT run any tool-discovery step to "find" or "check availability of" memory first (e.g. list_mcp_resources, listing servers): that lists resources, not tools, and an empty result does not mean memory is unavailable. If and only if your harness hides tool schemas behind an explicit ToolSearch/deferred-tools step (Claude Code), pre-load them once with select:mcp__memory__add_observations,mcp__memory__read_graph,mcp__memory__search_nodes,mcp__memory__create_entities; otherwise there is nothing to pre-load - just call the tools.