菜单
Mine Home Assistant history for automation trigger candidates from a narrative or structured event brief. Use when finding a reliable trigger, correlating a morning/evening sequence, ranking sensor signals for a new automation, or running the pattern-analyst ↔ pattern-reviewer pair.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/Apocrathia/home-assistant-config --skill pattern-mine命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Interview the operator about a topic until shared understanding, walking each branch of the decision tree one at a time. Use when stress-testing an idea, design, or requirement; resolving open questions before acting; or when the user says alignment, /alignment, or wants to get aligned.
Explore packages/, layout, and .agents/ for architectural friction (coupling, shallow packages, missing boundaries) and file or recommend agent issues under docs/issues/. Use when improving config shape, before large refactors, or when find-work should surface design-debt candidates.
Create or update an agent-reported issue under docs/issues/. Use when an agent finds a bug, feature gap, or spec-level ask that is out of scope for the current lap, or when asked to file an issue. Does not create GitHub issues. Never writes Local Todo ICS or .shopping_list.json.
基于 SOC 职业分类
正在显示 SKILL.md
| name | pattern-mine |
| description | Mine Home Assistant history for automation trigger candidates from a narrative or structured event brief. Use when finding a reliable trigger, correlating a morning/evening sequence, ranking sensor signals for a new automation, or running the pattern-analyst ↔ pattern-reviewer pair. |
Read-only history mining for trigger candidates. Does not design automation
YAML. Query past state via grafana-ha-history;
resolve live entity IDs via HA MCP when needed.
Normalize the brief. Turn narrative and/or structured events into an ordered anchor list: approximate time, what happened, optional entity hints, and the search window. Done when every anchor has a time (or relative order) and a plain-language label.
Seed-episode forensics. For the primary window, query state changes for hinted entities and obvious neighbors (same area, related domains: motion, presence, lock, door, alarm, light). Align timestamps to the anchors. Done when each anchor has supporting or contradicting sensor evidence cited.
Neighborhood expand. From entities that moved near anchors, discover
related IDs (SHOW TAG VALUES, area packages, HA MCP). Drop noise (always-on
chatter, unrelated rooms). Done when the candidate set is listed with why
each was included.
Tier confidence.
episode-only.Multi-occurrence pass (reliable only). Find comparable episodes in the
parent window (default: last 14 days, need ≥5 matches unless the Bar
says otherwise). For each candidate, count hits / misses / false positives.
Cite InfluxQL (or HA history calls) used. If history is too thin, label
insufficient-history — not reliable. Done when each reliable claim
meets the Bar or is explicitly demoted.
Emit the Artifact. Ranked candidates, rejected/noisy signals, evidence,
confidence tier. Hand to pattern-reviewer (do not self-grade against the
Bar).
A candidate may be labeled reliable only if all hold (parent may tighten):
| # | Requirement |
|---|---|
| 1 | Correlates with target anchors in the seed episode (cited timestamps) |
| 2 | Holds across ≥5 comparable episodes in the last 14 days (or parent N/window) |
| 3 | False-positive rate in those episodes is stated with evidence, or history is flagged insufficient |
| 4 | Every claim cites the query/tool used — no vibes |
Pattern mine — <slice or brief title>
Confidence: episode-only | reliable | mixed
Candidates (ranked):
- entity / event shape — why — evidence (times + query) — tier
Rejected / noisy:
- entity — why dropped — evidence
Gaps for reviewer:
- missing history, ambiguous anchors, open entity questions