monkey-skills

Append-only bitemporal store for validated operational-KPI series-points (US SEC primary-source layer). Persists file-per-series JSON keyed by company+kpi_id under a durable DATA dir, keeping full history so a restatement appends a superseding record and a point-in-time query sees only what was known then. 営業 KPI 二時制追加専用ストア。營運 KPI 雙時態 唯附加儲存層。

Layer-1 unified data fetch across US/JP/TW/KR/CN equities + macro — one pack.py facade, auto market detection from ticker suffix (.TW/.KS/.SS/.T/ bare-4-digit), 7 pack types (snapshot / memo-fetch / comps-multiples / screener-batch / regime-pack / kpi-quarterly / kpi-topline-backfill). Emits a raw data pack（原始資料包，非渲染卡片）— structured JSON straight from source clients including SEC EDGAR, EDINET, TWSE, and FRED (+14 more) through a shared cache layer; use this for a 資料層 health check, verifying cache writes/hits (快取), or fetching by source name (EDINET/EDGAR/TWSE/etc). Pure I/O, no analysis — for regime classification/判斷 (e.g. an Investment Clock verdict), use analysis-macro-regime instead. Consolidates the per-country data-{us,jp,tw,kr,cn} client scripts behind one CLI + shared cache layer.

Deliver a Mental-Model-first briefing before asking the user to decide a non-trivial engineering fork — the default, not optional. Also fires reactively when they're lost on the question, the explanation, or the stakes — including a 2nd consecutive check-question, which trips the repeated-confusion guard back to the Mental Model.

Family entry router for loom-discovery, the loom problem-space station. Use when unsure which discovery skill applies, or when a product idea's target users are still unclear (說不清是給誰用的 / not sure who this is for / 誰のためかまだ不明) — comes BEFORE product-principles. Routes worth-doing asks (值不值得做 / worth my time budget) to business-value, and user-needs asks (需求研究 / user research / ユーザーインサイト) to user-insights. Non-discovery asks redirect to their family entries. Members share no artifact and no agent.

Router for loom-interface-design — design a product's interface/interaction/UX in any modality (GUI/TUI/CLI). Routes to design-system (→ DESIGN.md) + interaction-flows (→ ui-flows.md). Use for 'design the UI/UX', screen flow, TUI/CLI layout.

The loom-product-principles family entry: intake + routing. Start here on a new product/feature idea when 不確定從哪開始 / 'where do I start' — checks the on-ramp criteria, then hands off to the member skill.

Use when starting spec work and unsure which loom-spec skill applies. Family entry router — routes to spec-expansion (draft/expand a spec from a seed) or completeness-critic (critique an existing draft for omissions).

Layer-3 read-only Markdown tearsheet for one company's recorded operational KPIs — periods as columns (newest left), one row per kpi_id, a †-marked disagreement cell + Revisions footnote when a later filing changed a value. Subprocesses kpi_store.py `dump` then a pure formatter; no fetch, no analysis, no verdict. Fires on "tearsheet", "KPI tearsheet", "KPI 一頁總覽", "KPI 總覽", "有沒有變", "KPI ワンページ", "何を記録した". KPI 觀察歷史一頁總覽。 KPI 観測履歴ワンページサマリー。

Re-distill the stale knowledge pages rescan flagged — LLM rewrites entities/metrics/concepts whose derived_from evidence changed, grouped by domain, skipping mature (human-certified) pages. Use for 'redistill', 're-distill stale knowledge', 'knowledge pages are stale', '重蒸餾知識頁'. This is the LLM half; cheap evidence-only refresh→rescan; evidence+knowledge in one shot→update; setup→init.

Evidence-layer-only refresh (advanced, 0 LLM): re-scan .dbt-wiki/ evidence — diff manifest md5, reprocess only changed resources, re-run sqlglot lineage, preserve User Notes, flag stale knowledge pages. Use ONLY when the ask is explicitly the cheap evidence half: 'rescan dbt-wiki', 'evidence only, 0 LLM', '只更新證據層', 'エビデンス層だけ更新'. A plain maintenance ask ('model 改了', 'update the wiki', 'bring the wiki up to date', '更新 wiki') is NOT this skill — it is update, which runs this plus the phantom-column gate and the review handoff. Only the LLM half→redistill; setup→init; add notes→ingest.

THE maintenance verb for dbt-wiki — one full pass: (optional) ingest new tribal knowledge → rescan evidence → gated redistill of stale knowledge → phantom-column lint gate → review handoff → scorecard. Runs every mechanical step, then hands the human-judgment part to review. Use for 'update dbt-wiki', 'update the wiki', 'bring wiki up to date', 'maintain the wiki', 'sync dbt-wiki', 'model 改了 update everything', '同步 wiki', '更新 wiki', 'wiki 維護'. Cheap evidence-only→rescan; just re-distill stale→redistill; certify pages→review; setup→init.

Family-entry router for dbt-wiki — the knowledge base distilled from a dbt project (manifest + compiled SQL + column lineage) into .dbt-wiki/. Use when unsure which dbt-wiki skill applies: first-time setup, bringing the wiki up to date, asking a dbt question, or certifying pages. Triggers: 'dbt wiki', 'dbt 知識庫', 'dbt ドキュメント', '不知道該用哪個 dbt-wiki', 'どの dbt-wiki を使えばいい'. Do NOT use for an Obsidian vault wiki (obsidian:using-obsidian) or a source-repo wiki (repo-wiki).

Make defensible investment decisions from primary-source frameworks. Use for Buy/Hold/Sell verdicts, equity memos, portfolio rebalancing, macro regime calls (Investment Clock), or TW equity analysis (三大法人). Business strategy → research-team.

Layer-3 orchestrator for a full equity investment memo (Markdown). Routes by ticker to data-markets memo-fetch (market auto-detected), runs analysis-dcf + macro-regime, then delegates the memo protocol + gates to domain-teams:investing-team. Pure orchestration.

Run 15 health checks on wiki/ (structural, semantic, provenance). Use weekly or after wiki-ingest batches. Read-only — does not auto-fix. Do NOT use to fix issues (use wiki-ingest / wiki-cross-linker). Obsidian wiki 健全性チェック・健康檢查。

One-verb wiki maintenance orchestrator — delegates wiki-ingest → wiki-cross-linker, runs the mechanical fix loop (safe-tier repairs on a proposal branch), triages work orders, reports a structural scorecard. Use for "update my wiki" / 整理 wiki / wiki 維護 / 維護 wiki / ウィキ更新・ウィキメンテナンス. Do NOT use for a single-purpose run — new-note ingest → wiki-ingest; link strengthening → wiki-cross-linker; read-only health check → wiki-lint; near-dup pair merge → wiki-merge; questions → wiki-query; research gap-fill → wiki-auto-research; first-time scaffold → wiki-setup.

Pure compute 3-stage DCF on pre-fetched data JSON. Input: --input <pack-json-path> from data-markets/scripts/pack.py --pack memo-fetch. Output: intrinsic value range JSON with 3×3 sensitivity (WACC × terminal growth). DCF 内在価値純計算層。DCF 內在價值純計算層。

Use when a task takes >1 hour OR touches >1 module — splits work into atomic one-failing-test units, three subagents each (implementer / spec-reviewer / code-quality-reviewer).

Record, recall, or prune the repo-native loom practice-memory store (`docs/loom/memory/`) — save a distilled lesson/gotcha, check prior experience before loom work, or propose keep/merge/retire cleanup (never auto-deletes). CONDITIONAL: fires only when the repo has `docs/loom/memory/README.md` — otherwise `loom-memory: N/A` with the reason, loudly. NOT for git commit trailers (dev-workflow:git-memory) nor Claude auto-memory. Triggers: "有沒有相關經驗", "記住這個做法", "記憶整理", "この教訓を残して", "過去の知見はある？".

Use when ready to close out a development branch — about to merge or open a PR. Fires on 'finish this branch', 'wrap up', 'ready to merge', 'open a PR', 'ship it'. Orchestrates review → verification → git-memory commit → git push. No auto-merge.

Use BEFORE any push/merge/PR on a non-trivial branch — whole-branch review of the cumulative diff. Fires on 'review my branch', 'ready to merge?', and excuses it refuses: 'just push', 'skip review', git push / gh pr create with no prior review-PASS.

Use BEFORE declaring 'done' — claiming complete without running the package-level test suite. Fires on 'I'm done', 'ready to ship', 'tests pass' (no invocation shown), 'merge'.

Mandatory gate before every git commit / gh pr create / gh pr merge — the skill decides whether memory trailers (Decision/Learning/Gotcha) apply; don't pre-judge a commit 'routine'. Also recalls past decisions: 'why did we…', '為什麼', an old branch.

Router for loom-code — invoke when the user wants to build, change, debug, or review code (features / fixes / refactors / migrations / reviews). Drives brainstorm → plan → SDD → TDD → debugging → code review → finish branch.

Use AFTER brainstorming produces a brief, BEFORE subagent-driven-development dispatches implementers. Splits it into atomic tasks — each with one RED/GREEN test and a single module boundary — into a dependency graph.

Adversarial worth-it check before betting real time on a product-shaped idea. Use for worth it? / should I build this / weighing ideas against one time budget. Market sizing → domain-teams:planning-team. 值不值得做 / ビジネスバリュー / 時間の使い方.

The problem-space research verb of loom-discovery — map, with recorded evidence, WHAT users need and which needs are worth serving, before any design or spec work. Two modes: opportunity-space mapping (knowledge work — research the world, never interrogate the user for researchable facts) and value commitment (a value judgment — the agent proposes a recommendation and writes the commitment only after the user ratifies it). Produces user-insights.md + evidence.md + per-question research/ reports under docs/loom/discovery/<date>-<slug>/. Problem-space-pure: states needs and outcomes, never solutions. Delegates heavyweight research to research-toolkit:deep-deep-research. Use for: user needs / what do users need / needs research / 使用者需求 / 使用者洞察 / ユーザーインサイト. Not for business worth-it verdicts (that is business-value) nor for designing the solution (that is loom-interface-design / loom-spec).

Adversarial critique of a design output (DESIGN.md + ui-flows.md) via a writer≠judge panel for SURFACE omissions (undrawn states, dead-ends, a11y) per Nielsen. Use to review a design draft for gaps before spec-expansion. Surface only, never code.

Drives the loom principles→design→spec→code pipeline end-to-end via deterministic Claude Code Workflow segments. Use when asked to run the whole pipeline or to auto-implement a change from principles through code. CONDITIONAL: fires only when the Workflow tool is available AND the four station plugins are installed — otherwise `loom-pipeline: N/A` with the reason, loudly (Codex hosts are N/A). Triggers: "run the loom pipeline", "全管線", "全流程跑一遍", 自動實作, "run the conductor".

Adversarial completeness critique of a spec-expansion draft via a critic panel hunting OMISSIONS (missing objects/actors, unhandled states, NFR gaps). Use to review a spec draft for gaps before VERIFY. Spec omissions only, never code.

Turn a sparse seed (a few lines of feature intent) into a high-recall spec draft — fan out objects, states, paths, edge cases → acceptance criteria in OpenSpec shape. Use for requirement fan-out / edge-case coverage before implementation.

Generate a product's visual design system — colors, typography, layout, component tokens — per PRINCIPLES.md. GUI → DESIGN.md (YAML tokens + prose); TUI/CLI → conventions stub. Use BEFORE spec/build to define the look. Flows → interaction-flows.

Generate a ui-flows.md interaction-flow artifact — screens, navigation, user flows, state transitions, wireframes (GUI/TUI/CLI). Use when the user wants to map, design, or lay out a UI/UX flow, screen, or its states before specing or coding.

Turn a sparse product idea into a PRINCIPLES.md constitution (north star + falsifiable principles). Use BEFORE design/spec/build, or when asked what principles should guide a product/design/engineering decision or trade-off. Design/spec critique → design-critic/completeness-critic. Triggers: 產品原則 / 設計原則 / 工程原則 / 產品憲章 / プロダクト指針.

Build or modify domain-team skills with convention discipline — primary-source grounding, 3-commit split, 4-tier gates. Use when creating or refactoring a team skill. Generic Claude skills → superpowers:writing-skills.

Score Agent Skill design quality on an 8-dimension rubric (0-120 + grade A-F). Use when reviewing or scoring a SKILL.md / skill package, or 'is this skill well-designed?'. Behavioral testing → dogfood-skill-testing; redesign → skill-creator-advance.

Use BEFORE implementing new behavior or non-obvious design — explore intent + alternatives via an upstream-artifact gate (Axis 0) → a brief. Refuses 'this is simple' / 'just start coding'.

Use when the branch touched UI and ui-flows.md exists — drive each state before branch close, otherwise N/A. GUI only; not for package tests (verification-before-completion). Triggers: "verify the UI", "did the screens actually render", 畫面驗證, UI 検証.

Create new skills, redesign (phases, agents, workflow/contract), run evals, or optimize a description's triggering. Use for 'build a skill', 'make a slash command', or a redesign. For output A/B use skill-tuning; token refactor use skill-refactor.

Pure-compute cross-validation of US SEC financial-statement doc-table cells against XBRL companyfacts. Input: --source-a <doc-table-cells-pack> --source-b <companyfacts-pack>. Output: comparisons + high_alerts + single_source JSON, classifying each finding two-source vs single-source (scale/rounding, DQC 2.4.1 decimal-disagreement, restatement-signal). Xval 財報主表逐格對 XBRL 交叉驗證純計算層。