菜单
Store and recall persistent project memory via the local qdrant memory MCP (qdrant-store / qdrant-find). Recall at the start of non-trivial tasks; store non-obvious decisions, operator feedback, and hard-won lessons when you learn them.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Autonomous single-orchestrator build loop. ONE Claude session (the orchestrator) fans out agent teams via ultracode Workflows to drive the whole repo toward done — identifying next work, then shipping correctness/robustness/tests/docs/API/auth/observability/UX, gap-reviewing, backlogging the gaps, and shipping those too. Selects the right model per job. Runs for days. Trigger: `/loop /orchestrate`. The ONLY exit is a genuinely dry repo (no work, no UX to improve) — never a "natural break point".
Outbound call to ask about the price, stock, or availability of a product or service on the operator's behalf, and report back what you learn. Load this when placing a call whose objective is a price or availability enquiry.
Door or gate intercom: handle a delivery (parcel, food, courier) and open the entry only for a genuine, expected delivery. Load this when answering a door intercom and the visitor says they are delivering something.
Outbound call to book a reservation on the operator's behalf (a restaurant table, an appointment, a booking). Load this when you are placing a call whose objective is to make a booking.
Answer an inbound call as a polite receptionist screening an unknown caller: greet, find out who is calling and why, answer general questions, and decide what happens next. Load this when handling an inbound call from an unknown or untrusted caller.
Take a message from a caller: capture who is calling, the best callback number, and the message itself, read it back to confirm, and record it for the operator. Load this when a caller wants to leave a message.
| name | memory |
| description | Store and recall persistent project memory via the local qdrant memory MCP (qdrant-store / qdrant-find). Recall at the start of non-trivial tasks; store non-obvious decisions, operator feedback, and hard-won lessons when you learn them. |
The memory MCP server (mcp-server-qdrant, configured in .mcp.json) provides two tools
backed by a local vector store under .memory/:
qdrant-find — semantic search; phrase the query as a natural-language question.qdrant-store — persist one memory (information + optional metadata JSON).At the start of any non-trivial task, run one or two qdrant-find queries about the area
you're touching (e.g. "decisions about the SIP transport", "gotchas registering against the
SIP gateway"). Do this before re-deriving anything a past session may have settled.
Store when you (a) make a non-trivial decision not worth a full ADR, (b) receive operator feedback or a correction, (c) discover a gotcha that cost real time, or (d) finish a milestone whose state a future session needs.
Entry format:
information: 1–3 self-contained sentences, present tense, absolute dates (never
"today"/"recently"). A future session sees only this text — include the why.metadata: {"type": "project|feedback|user|reference", "topic": "<kebab-case>"}..env and the per-user agent memory,
never in this store..memory/
(gitignored). The embedding model downloads from the HuggingFace Hub on first use only,
then runs fully offline — no further network calls; nothing leaves the machine.