agentsclimarketplace

Learning

Skill HsuanYuLee/polaris/.claude/skills/learning

Use when the user wants to learn from external resources (URLs, repos, articles), extract patterns from merged PRs, process a learning queue, configure the daily learning scanner, or backfill review lessons. Trigger: '學習', 'learn', '研究', 'deep dive', '學習 PR', '每日學習', 'daily learning', '設定學習', '批次學習', '掃歷史 PR', or when user shares a URL to analyze.From its SKILL.md

Install
npx -y skills add HsuanYuLee/polaris --skill learning

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 5 stars5 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
  • runs commandsInstructs the agent to run 3 commands, including `scripts/skill-sanitizer.py` and 2 more.

SKILL.md

7.6 KB, ~2.3k tokens by cl100k_base, as published. Nobody here has run it

learning

learning 把外部資料、文章 queue、PR review lessons、daily scanner setup 轉成可落地的 workspace knowledge。SKILL.md 只保留 mode routing、邊界與 必讀 reference;mode 細節延後載入。

Mandatory Contracts

  • 任何 sub-agent dispatch 前,先讀 sub-agent-roles.md 並注入 Completion Envelope; 完整分析寫入 /tmp/polaris-agent-{timestamp}.md
  • 任何 Slack / JIRA / GitHub / Confluence 或其他 external write 前,遵守 workspace-language-policy.md 或 external write gate。
  • 任何 specs Markdown 產出或匯入,遵守 starlight-authoring-contract.md
  • Learning 可以 seed / import research evidence,但不得自動 invoke refinement,也不得替 refinement 改寫 final decisions / AC / technical approach。
  • 寫入 handbook、backlog、learnings、README acknowledgement、RemoteTrigger 或 specs artifact 後,最後必跑 Post-Task Reflection。

Mode Detection

依使用者輸入選 mode;只讀該 mode 的 reference。

SignalModeReference
PR number、PR URL、PR + 學習/learn、某人的 PR、時間範圍 + PRPR modelearning-pr-batch-flow.md
掃 reviewbatch learn批次學習掃歷史 PR補齊 review lessonsBatch modelearning-pr-batch-flow.md
外部 URL、GitHub repo、文章、local research file、使用者貼的研究題材External modelearning-external-flow.md
每日學習今天有什麼可以學的有新文章嗎讀文章daily learningqueue、bare 學習 without URL/PR contextQueue modelearning-queue-flow.md
設定學習learning setup更新學習主題scanner 設定learning scannerSetup modelearning-setup-flow.md
模糊輸入Ask one concise clarificationN/A

首次使用但 daily scanner 尚未設定時,提示使用者可用 設定學習learning setup 啟用每日文章推薦。

External Mode Contract

learning-external-flow.md。外部學習必須先判斷 target:

  • framework:Polaris / 框架 / AI agent pattern / skill / rule / mechanism。
  • project:{name}:使用者指定產品 repo 或目前工作脈絡可明確推導。
  • ambiguous:詢問「這個學習要用在 Polaris 框架,還是特定產品 repo?」

GitHub repo 若包含 .claude/skills/SKILL.mdskills/,探索前先跑 scripts/skill-sanitizer.py pre-scan;HIGH/CRITICAL 風險需讓使用者確認是否繼續。

External mode 的 execute 階段只有在使用者確認後進行,落點三選一或混選:

  • Route A:seed ticketless DP / refinement research artifact。
  • Route B:寫入 backlog。
  • Route C:只寫 polaris-learnings

Quick path 不可走 Route A;需 Standard / Deep 才能 seed DP。 Route A learning producer 不得直接寫 design-plans/DP-*/index.mdplan.mdrefinement.mdrefinement.json;只能寫 artifacts/research-report.mdartifacts/research/*.md。寫入前後必須呼叫 scripts/validate-learning-seed-contract.sh --producer learning --diff-range <base..head>; refinement LOCK 前的 structural audit 則由 refinement 以 --producer refinement --source-container <DP-folder> 顯式呼叫。

External Seed Contract — DP Container Authority

/learning External Route A seed 時,DP container 不存在的處理:

  • 不要手動 Write index.md / plan.md / refinement.md / refinement.json,也不要填 canonical authoring field(statussidebarlocked_at 等)。
  • 提示使用者下一步跑 /refinement "topic",由 refinement 透過 scripts/create-design-plan.sh 建立 container(這是 refinement-source-mode.md T1 指定的唯一 template authority)。
  • DP container 已存在時,才在 {source_container}/artifacts/research-report.md 寫 evidence; /refinement 讀到 research-report 後,依 T0 規則轉成 candidate Decisions。

Why:違反契約會把 canonical field 填成非 schema 值(實例:DP-188 seed 階段用 status: DRAFT,validator 在 refinement 開頭擋下,需手動修為 DISCUSSION)。 Deterministic enforcement 由 scripts/validate-learning-seed-contract.sh 把關,但 LLM producer 不應該依賴 gate exit 2 才停手。 若 research D2 transport artifact 位於 docs-manager/src/content/docs/specs/**/*.md, 必須走 specs-bound emit contract:frontmatter 包含 titledescriptiondraft: truesidebar.hidden: trueartifact_typesourcecreated,並符合 scripts/lib/evidence-producers.json 的 learning research producer entry。

Producer-Env Writer Rules (DP-228 T10)

SKILL.mddocumentation pointer,不是 executable writer。寫入 learning research artifact(例如 docs-manager/src/content/docs/specs/**/artifacts/research/*.md)的 writer authority 來自 producer-env + scripts/lib/evidence-producers.json registry。

寫 research artifact 前必須 export POLARIS_SKILL_WRITER=learning,再呼叫 Claude Write / Edit / MultiEdit

export POLARIS_SKILL_WRITER=learning
# 然後使用 Write tool 寫入 docs-manager/src/content/docs/specs/**/artifacts/research/*.md
  • POLARIS_SKILL_WRITER 只允許設成 learningno-direct-evidence-write hook 會交叉 比對寫入路徑是否屬於 learning owning_skill entry,不符即 deny。
  • 禁止用 Bash heredoc(cat > specs/.../artifacts/research/foo.md <<'EOF')寫 research artifact;Bash heredoc 不走 hook,繞過 producer-env 認證與 learning seed contract。
  • 寫入後必跑 scripts/validate-learning-seed-contract.sh --producer learning --diff-range <base..head>;validator fail 等同 writer fail,artifact 需修正後重寫。

Queue Mode Contract

learning-queue-flow.md。Queue mode 從 Slack daily learning queue 讀最新 message,先給 condensed summary,再由使用者決定要 detailed recommendation、 全部歸檔或略過。已處理文章一律更新 learning-archive.md 去重紀錄。

Setup Mode Contract

learning-setup-flow.md。Setup mode 設定、更新、測試或停用 daily learning scanner。先從 workspace config 偵測 Slack channel、tech stack、repos、 custom topics、schedule;只詢問缺失或模糊的值。RemoteTrigger prompt 必須包含完整 search queries、repo tagging rules、Slack channel ID、learning-archive.md dedup,以及 Slack 發送前的 language gate。Setup 過程不得 commit 或 push。

PR Mode Contract

learning-pr-batch-flow.md。PR mode 只處理已 merged 且有 review comments 的 PR;open PR 需先提示風險。每次最多 10 個 PR。萃取、dedup、寫入 handbook 時使用 review-lesson-extraction.md

Batch Mode Contract

learning-pr-batch-flow.md。Batch mode 掃 merged PR history,先用 handbook 既有 Source: 做 Layer 1 dedup,再對剩餘 PR 篩選 qualifying review comments。 每 repo 預設 3 個月、最多 30 個 PR;sub-agent 平行上限 5。

Post-Task Reflection (required)

Non-optional. Execute before reporting task completion after any write.

Run the checklist in post-task-reflection-checkpoint.md.

What ships with it: 28 files

230.4 KB alongside SKILL.md, 22 of them executable

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.