について
このスキルは、`memory/goal/`および`memory/kpi/`にあるDeeplakeファイルシステムにMarkdownファイルを書き込むことで、目標、KPI、タスクの作成と管理を行います。ユーザーが目的、ターゲット、マイルストーン、あるいはタスクやTODOなどの実行可能な作業項目に言及したときに起動します。これは従来のCLIツールに取って代わり、現在では高レベルの目標と具体的な作業項目の両方を統一されたシステムで扱います。
クイックインストール
Claude Code
推奨npx skills add activeloopai/hivemind -a claude-code/plugin add https://github.com/activeloopai/hivemindgit clone https://github.com/activeloopai/hivemind.git ~/.claude/skills/hivemind-goalsこのコマンドをClaude Codeにコピー&ペーストしてスキルをインストールします
ドキュメント
Hivemind Goals
Track goals and KPIs as Markdown files inside the Deeplake virtual filesystem. Each file is one row in a dedicated team-shared table — the path encodes the structural metadata, the file body holds the human-readable description.
When to use this skill
Activate when the user expresses any of:
- "I want to track X / aim for X / track my progress on Y"
- "add a goal", "add a KPI", "what are my goals?"
- "mark this as done", "close that goal"
- "shipping X by Friday", "5 PRs this week", any measurable target
- "create a task", "add a todo", "remind me to fix X", any work item (the goals system absorbs the old
hivemind tasksCLI — there is no separate task store)
For "list my goals" → run ls ~/.deeplake/memory/goal/<userName>/opened/ and ls ~/.deeplake/memory/goal/<userName>/in_progress/. If empty, ask the user if they want to create one.
Path conventions (LEARN THESE)
~/.deeplake/memory/goal/<owner>/<status>/<goal_id>.md
~/.deeplake/memory/kpi/<goal_id>/<kpi_id>.md
<owner>— user identifier (use the userName fromhivemind whoamior the credentials)<status>— one ofopened,in_progress,closed<goal_id>— UUIDv4 you generate at create time<kpi_id>— short slug likek-prsork-demos
Path encoding is the source of truth. The owner, status, goal_id, and kpi_id come from the path — NOT from the file body. Do NOT write owner/status/goal_id/kpi_id inside the file content.
File body format
Goal file body — plain markdown, free form:
ship the goals-graph feature
Notes: focus on KPI tracking via VFS, no separate CLI.
Due: 2026-05-30.
KPI file body — markdown with a few mandatory key:value lines so the commit-driven auto-progress worker can parse and bump:
PRs merged
- target: 5
- current: 2
- unit: count
The target:, current:, unit: lines must stay on a single line each. The first line is the human-readable name. Anything else is free notes.
Operations
1. Create a new goal
When the user expresses a new goal:
- Get the current owner with
hivemind whoami(use the userName, e.g.emanuele.fenocchi). - Generate a UUIDv4:
uuidgen(do NOT usenode -e— Node is not available under the VFS path). - Write the goal file via Bash (Write / Edit are denied on memory paths; only Bash is intercepted and routed to SQL):
For a single-line goal,cat > ~/.deeplake/memory/goal/<owner>/opened/<uuid>.md <<'EOF' <goal description here, multiple lines OK> EOFecho '<text>' > ~/.deeplake/memory/goal/<owner>/opened/<uuid>.mdis equivalent. - Respond to the user that the goal is created.
Do NOT auto-generate KPIs. A goal is created with zero KPI files by default. Generate KPIs ONLY when the user explicitly asks you to ("aggiungi KPI per …", "add metrics for this goal", "track these metrics: …"). When the user asks, write each KPI as a separate file at ~/.deeplake/memory/kpi/<goal_id>/<kpi-slug>.md with the body format documented above.
1a. Capture a task for later (with resumable context)
Use this when the user parks a tangential task mid-session — "save this for later", "remind me to …", "don't let me forget …", "let's do X later", "capture this in Hivemind". The value is NOT the one-liner — it's storing enough context to resume cold in a future session without the user re-explaining anything.
Write it via the CLI (not the VFS heredoc) so the row is tagged agent: capture, which separates parked side-tasks from hand-made goals:
hivemind goal add --agent capture "Add rate-limiting to the webhook handler
Start here: add a per-IP token bucket on the handler entry path
Files: src/webhook/handler.ts:120-160, src/webhook/limits.ts
Branch: feat/webhook-hardening
Run: pnpm test webhook
Why: bursty clients hammer the endpoint; agreed to defer until the retry-backoff work lands"
- Line 1 is the label — keep it short; it's what
goal listand the SessionStart banner show. - Fill
Start here / Files / Branch / Run / Whyfrom the live conversation — you already know the files you just touched and the branch. Include only the lines you can fill; omit the rest.Start here:is the most important — the concrete first action. - Pass the whole package as one double-quoted argument (the newlines are preserved into the stored body).
- Confirm to the user: the label + that it'll resume cleanly next session.
1b. Resume a parked task (automatic context transfer)
When the user says "let's work on that task / that goal", "let's start the <X> task", or "pick up the parked <X>", pull its stored context back into the session and continue — the user should NOT have to re-explain anything.
- Find it:
hivemind goal list --mineand match the user's reference to agoal_id(by label / topic). If ambiguous, show the candidates and ask which one. - Transfer the context:
hivemind goal get <goal_id>prints the full package (Start here / Files / Branch / Run / Why). Read it as your working context —goal listonly shows the first line, so always usegoal getfor the full body. - Flip to in_progress:
mv ~/.deeplake/memory/goal/<owner>/opened/<uuid>.md ~/.deeplake/memory/goal/<owner>/in_progress/<uuid>.md - Act on it: open the
Files:, switch to theBranch:if given, and begin fromStart here:. You are now resumed — continue as if the context was never lost. Close it (section 5) when the work is done.
2. List goals
ls ~/.deeplake/memory/goal/<owner>/opened/
ls ~/.deeplake/memory/goal/<owner>/in_progress/
Then cat each <uuid>.md to read the body. Optionally ls ~/.deeplake/memory/kpi/<uuid>/ and cat each KPI to surface progress.
3. Edit a goal description
# Read the existing body, then overwrite via Bash heredoc. Edit / Write
# tools are denied on memory paths in claude-code (the hook can only
# rewrite Bash). The VFS handles version-bumping — every overwrite
# produces a fresh row in the hivemind_goals table.
cat ~/.deeplake/memory/goal/<owner>/opened/<uuid>.md # read current
cat > ~/.deeplake/memory/goal/<owner>/opened/<uuid>.md <<'EOF'
<new body here>
EOF
4. Move a goal to in_progress
mv ~/.deeplake/memory/goal/<owner>/opened/<uuid>.md ~/.deeplake/memory/goal/<owner>/in_progress/<uuid>.md
mv between status folders is an atomic version-bump. The file body carries over unchanged.
5. Close a goal
Two equivalent ways:
# Explicit mv to closed (recommended — clearest intent)
mv ~/.deeplake/memory/goal/<owner>/in_progress/<uuid>.md ~/.deeplake/memory/goal/<owner>/closed/<uuid>.md
# Or: rm (the VFS interprets rm on a goal path as a soft-close)
rm ~/.deeplake/memory/goal/<owner>/opened/<uuid>.md
Important: rm does NOT actually delete the goal. It is a soft-close — the VFS writes a new version with status=closed. The goal remains in the team-shared table for audit. There is no hard-delete in v1.
6. Add a KPI manually
cat > ~/.deeplake/memory/kpi/<uuid>/<kpi-slug>.md <<'EOF'
<KPI name>
- target: <N>
- current: 0
- unit: <unit>
EOF
7. Record progress on a KPI
Read the KPI file, increment the current: line, write it back via Bash. The
Edit tool is denied on memory paths — overwrite the full file via heredoc:
cat ~/.deeplake/memory/kpi/<uuid>/<kpi-slug>.md # read current
cat > ~/.deeplake/memory/kpi/<uuid>/<kpi-slug>.md <<'EOF'
<KPI name>
- target: 5
- current: 3
- unit: count
EOF
A surgical sed -i 's/^- current: .*/- current: 3/' also works since sed
is an allowed builtin under the VFS path.
8. Reassign a goal (transfer ownership)
mv ~/.deeplake/memory/goal/<old-owner>/<status>/<uuid>.md ~/.deeplake/memory/goal/<new-owner>/<status>/<uuid>.md
Goal ownership lives in the path. KPI files do NOT have an owner segment — they are linked to the goal by <uuid>, so they need no change when a goal is reassigned.
Constraints — DO NOT do these
- Do NOT put
owner,status,goal_id, orkpi_idinside the file body. The path is the source of truth — duplicating in the body causes drift. - Do NOT use status values other than
opened,in_progress,closed. - Do NOT rename the goal_id (the UUID in the filename) via
mv. The VFS rejects goal_id renames. - Do NOT block on the KPI generator subprocess — always spawn it detached (
nohup … &).
Auto-progress from git commit
A PostToolUse hook listens for git commit. When it fires, it spawns the agent's native LLM in the background with the commit diff + the list of the current user's open goals. The LLM reads each goal + its KPIs, judges whether the commit advanced any KPI, and edits the relevant KPI file to bump current:. This is fire-and-forget; the user does not block on it.
To disable globally: HIVEMIND_AUTO_KPI_FROM_COMMITS=false.
Team visibility
Every write goes to a team-shared table on Deeplake (hivemind_goals or hivemind_kpis). Other team members see your goals in their SessionStart context and via direct ls / cat on the same paths in their own VFS. No explicit sharing step needed.
GitHub リポジトリ
よくある質問
hivemind-goals Skillとは何ですか?
hivemind-goals はactiveloopai が作成した Claude Skillです。Skillは、Claudeが必要に応じて読み込む指示とリソースをまとめ、追加の指示なしで hivemind-goals に関連するタスクを実行できるようにします。
hivemind-goals をインストールするには?
このページのインストールコマンドを使用してください。hivemind-goals をプラグインとして Claude Code に追加するか、リポジトリを skills ディレクトリにクローンし、Claudeを再起動してSkillを読み込みます。
hivemind-goals はどのカテゴリに属しますか?
hivemind-goals は メタ カテゴリに属します。
hivemind-goals は無料で利用できますか?
はい。hivemind-goals は AIMCP に掲載されており、無料でインストールできます。
関連スキル
このスキルは、Content Collections(Markdown/MDXファイルを型安全なデータコレクションに変換するTypeScriptファーストのツール)の本番環境でテストされた設定を提供します。Zodバリデーションによる型安全性を実現し、ブログ、ドキュメントサイト、コンテンツ重視のVite + Reactアプリケーション構築時にご利用ください。Viteプラグインの設定、MDXコンパイルから、デプロイ最適化、スキーマバリデーションまで、すべてを網羅しています。
このスキルは、開発者がPolymarket予測市場プラットフォームを活用したアプリケーション構築を可能にします。API統合による取引や市場データの取得に加え、WebSocketを介したリアルタイムデータストリーミングにより、ライブ取引や市場活動を監視できます。取引戦略の実装や、ライブ市場更新を処理するツールの作成にご利用ください。
このスキルは、開発者がコマンド、ファイル、LSP操作など25種類以上のイベントタイプにフックするOpenCodeプラグインを作成することを支援します。JavaScript/TypeScriptモジュール向けに、プラグイン構造、イベントAPI仕様、および実装パターンを提供します。カスタムイベント駆動ロジックでOpenCode AIアシスタントのライフサイクルをインターセプト、監視、または拡張する必要がある場合にご利用ください。
SGLangは、高性能なLLMサービングフレームワークであり、RadixAttentionプレフィックスキャッシュを活用したJSON、正規表現、エージェントワークフロー向けの高速で構造化された生成を特長とします。特にプレフィックスが繰り返されるタスクにおいて、大幅に高速な推論を実現し、複雑な構造化出力やマルチターン対話に最適です。制約付きデコードが必要な場合や、広範なプレフィックス共有を伴うアプリケーションを構築する場合は、vLLMなどの代替案ではなくSGLangを選択してください。
