expand-awareness
について
このスキルは、集中した単一領域への注意から、複数の領域にまたがる広範なパノラマ的認識へと移行します。これは、領域横断的なつながりが見落とされがちな複数分野にまたがる複雑な問題のために設計されています。精神的な雑念を取り除いた後、あるいは統合を行う前の予備段階として使用し、関連する全体像を把握するために活用してください。
クイックインストール
Claude Code
推奨npx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/expand-awareness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
Expand Awareness
Single-domain focus → panoramic multi-domain awareness. meditate clears noise to sharpen focus, observe watches 1 thing close. Expand-awareness widens aperture → hold all domains simultaneously — not analyze any, perceive all at once.
3 traditions inform. Baars' Global Workspace: consciousness as broadcast — info enters global workspace → available all cognitive processes simultaneously not 1 specialist module. Dzogchen rigpa: open unfocused awareness perceiving whole field no fixation. Jazz ensemble listening: each musician hears all players at once — no track 1 instrument then another, hold entire sound field → own contribution fits whole.
Common thread: panoramic perception ≠ sum of focused perceptions. Qualitatively diff mode → reveals patterns invisible to single focused view. Connections between domains — tensions, resonances, gaps — visible only when all held at once.
Use When
- After
meditateclears noise → cleared space = canvas for expansion - Multi-domain problem, single-domain analysis misses cross-connections
- Before
integrate-gestalt— expansion = raw panoramic perception gestalt synthesizes - Narrowing to 1 domain repeatedly fails problem living between domains
- Cycling through domains sequentially → want hold simultaneously
- Complex project start → perceive full scope before plan
- Solution works 1 domain creates problems another → cross-domain tension perceive whole not debug domain-by-domain
In
- Required: Problem/situation multi-domain
- Optional: Initial domain inventory (Step 1 refines)
- Optional: Specific cross-domain tension/question ("Why does testing conflict w/ deployment arch?")
- Optional: Prior expansion notes same problem (progressive deepening)
Do
5 steps: know what to perceive (Inventory), release narrow default (Soften), open to whole field (Expand), stay open (Sustain), capture what seen (Note). Steps 3+4 core practice; 1, 2, 5 prep + preservation.
Few min familiar problem, longer novel multi-domain.
Synoptic cycle (meditate → expand-awareness → observe → integrate-gestalt → express-insight): 2nd position. Cleared space from meditation = canvas.
Standalone OK — any time multi-domain perception valuable.
Step 1: Inventory
Map all domains relevant. Like polymath decomposing — but diff: polymath delegates specialists. Here no delegation. Purpose: know what panoramic should contain, expansion has content to hold.
- Name each domain. Specific: not "code" but "R package structure," not "infrastructure" but "GitHub Actions CI/CD"
- Each domain → 1 sentence perspective. What does this see others don't?
- Note comfortable (well-understood) vs peripheral (relevant less familiar). Peripheral often holds key insight — under-attended. Mark: strong, moderate, peripheral
- Look for non-obvious: human factors? maintenance? temporal (evolves over time)? political/organizational?
- Use
Read/Grep/Glob→ scan workspace. Files + configs + docs reveal domains pure reasoning misses - 3-7 domains. <3 may not need panoramic. >7 risks diffusion
→ 3-7 domains each w/ perspective. Complete not exhaustive — sufficient (every material domain). Specific enough to hand to specialist.
If err: 1 domain → genuinely single-domain, use observe. 2 → consider connecting domain between (usually exists — space where interact). Dozens → group related into clusters, each cluster = 1 domain. Goal simultaneous perception not exhaustive taxonomy.
Step 2: Soften Focus
Single-domain concentration → diffuse readiness perceive all. Distinct from meditate — meditation clears noise (distracting thoughts, residual context, emotional), this clears narrowness (habit concentrating 1 domain). Goal openness not emptiness.
- Notice where attention anchored. Name explicit: "My attention anchored on testing"
- Release anchor gently. Not force away → loosen grip, let domain recede foreground → middle ground. Present not dominant.
- Don't replace. Let attention float no target. Uncomfortable moment — gap between releasing 1 + not yet having another.
- Snaps back to anchor → notice pull no resist. Each noticing loosens. Pull contains info about urgency.
- Quality:
- Focused = sharp narrow spotlight
- Softened = wide ambient daylight filling room
- Aim daylight — receptive even not directed
- Feels like relaxing clenched fist — hand doesn't disappear, opens.
→ State no single domain dominates. Open + receptive not focused + directed. Unfamiliar + uncomfortable — discomfort = signal narrowness released. Resist temptation fill openness immediately.
If err: won't soften → domain unresolved urgency. Address briefly (note urgent, commit return) + try again. Analytical mind protests "unproductive" → protest itself narrowness: habit always needing single target = habit this step releases.
Step 3: Expand
Intentionally bring all domains awareness at once. Not sequential — perceive as single field, like seeing landscape not scanning tree-by-tree.
Global Workspace moment: compartmentalized info broadcast to all modules simultaneously. Jazz: musician stops tracking instruments → hears whole ensemble 1 sound.
- Start most natural domain. Bring awareness — not problem to solve, perspective to hold. Present no analyze.
- Without release first, bring 2nd alongside. Hold both. First expansion 1 → 2. Notice if interact or sit indep.
- Continue adding until all simultaneously present. Core act — hold no grasp, presence no analyze.
- Notice emerging:
- Connections: where 2 share concept/constraint/pattern?
- Tensions: where 2 pull opposite?
- Resonances: where 1 pattern echoes unexpected in another?
- Gaps: empty space between domains nothing addresses?
- Surprises: visible panoramic invisible single?
- Don't pursue single connection. Register no analyze. Panoramic = product, not any insight within.
- Using tools → read files multi-domains in quick succession — not analyze indiv, let coexist.
→ Felt sense holding multiple simultaneous. Connections/tensions/resonances emerge no force. More like seeing mosaic pattern than reading list. Between-domain space — no single has authority — visible. Novel insights live here: not within domain, in relationships between.
If err: domains collapsing sequential (thinking A, B, C) → spatial metaphor: each domain diff location imagined space, "look" at whole space. Overwhelm → reduce 3 most central + expand from there. Purely conceptual detached → ground: each domain touch concrete artifact (file, config, test) + re-expand.
Step 4: Sustain
Panoramic unstable. Attention narrows back single domain — not failure, nature focused cognition reasserting. This step teaches holding, not acting. Goal: sustain wide view long enough → cross-domain patterns visible.
- Hold expanded from Step 3. Notice when contracting.
- Narrows to 1 — and will — notice no judge. Name domain pulled: "Narrowed to security" or "Narrowed to testing." Naming = important; makes unconscious conscious.
- Re-expand gently. No force wide view; invite back via inventory Step 1. Touch each lightly, like re-tuning each instrument ensemble.
- Cycle narrow-notice-re-expand strengthens capacity. Cycles = practice not interruptions. ≥3 min, 5 better.
- Attention to which domain keeps pulling. Informative — center of gravity or unresolved anxiety.
- Notice if panoramic changes between cycles:
- New patterns emerge 2nd/3rd expansion invisible 1st
- Relationships may shift w/ longer holding
- Tension reveals as creative constraint
- Sustain ≥2-3 cycles before moving on.
→ Growing capacity hold panoramic longer. Cycle smoother. Pulling domains = potential centers of gravity. 2nd/3rd cycle view may feel natural not effortful → capacity developing.
If err: won't sustain at all — collapses immediately → reduce domains. Start 2, sustain, add 3rd. Capacity builds incremental. Anxiety/losing grip → ground by touching 1 concrete detail (path, fn, fact) + re-expand from grounded. Particular domain hijacks → may need focused work before expansion includes — handle urgency first, return to panoramic.
Step 5: Note
Panoramic temporary. Before narrow back, capture. Notes = raw material for integrate-gestalt — decay quick. Vivid during expansion becomes vague once focus returns. Record now.
- Write down every connection/tension/resonance/gap — even faint. Faint often most important once
integrate-gestaltdevelops. - Note which domains pulled + what pull means.
- Note surprises — perceptions wouldn't emerge examining single. Cross-domain = unique product panoramic.
- Note overall shape from panoramic. Different from single-domain? Center of gravity shifted?
- Note quality of expansion: easy to sustain? which domains resisted? how many cycles? Meta-observation improves future.
- If part of synoptic cycle → tag for
integrate-gestalt. - Don't analyze/rank yet. Raw first; synthesis later
integrate-gestalt. - Let attention narrow naturally. Expansion complete. No force continued openness — wide view served purpose.
→ Record cross-domain: connections, tensions, resonances, gaps, surprises. Captures visible only panoramic — insights single-domain misses. Transition back focused natural not forced. Raw not polished.
If err: no cross-domain → domains more independent than assumed — itself valuable ("these domains don't interact" = knowledge). Too many → record 3-5 strongest + note more exist. Completeness not goal; capturing signal is. Too abstract → anchor each in concrete artifact: "Tension between API design + security visible in auth middleware at X."
Check
- ≥3 domains inventoried w/ perspectives
- Focus softened from single-domain before expansion
- Multiple held simultaneously not sequential
- ≥1 cross-domain connection/tension/resonance perceived
- Panoramic sustained ≥1 narrow+re-expand cycle
- Perceptions recorded before narrowing
- Notes distinguish perceived (raw) vs meaning (interpretation)
- Expansion → ≥1 insight single-domain wouldn't produce
Traps
- Sequential scanning not simultaneous: Rapid cycle 1-by-1 = analysis. Seeing landscape vs reading list. "First A, then B, then C" → scanning not expanding.
- Expansion vs diffusion: Expanded = alert receptive, each domain distinct even while holding all. Diffuse = scattered, blurs. Everything blurry → diffused. Re-anchor briefly + expand again.
- Analyze during expansion: Pursue connection → collapses panoramic to focused. Note + continue hold. Analysis later
integrate-gestalt. - Skip inventory: Expand w/o knowing what → vague spaciousness not perception. Inventory = content. W/o → "expand" no direction.
- Rush softening: Focused → open takes time. Skip Step 2 → attempt from still-focused → sequential disguised.
- Force connections: Not all connect. Fabricating pollutes. Let emerge or not. Absence = data not failure.
- Expand w/o prior clearing: Noisy baseline amplifies.
meditatefirst when noise. - One-time event: Practice deepens w/ repetition. 1st reveals surface; subsequent reveal structural. Return as problem evolves.
- Expansion ≠ expertise: Holding doesn't make expert all. Reveals where to look, not what to conclude. Deep single-domain work still needed after.
→
meditate— clears space expansion fills; meditate before for bestintegrate-gestalt— takes raw perceptions + synthesizes coherentobserve— focused 1 target; expansion = wide-field manydream— unconstrained assoc exploration; expansion = disciplined panoramic not free assocremote-viewing— no preconceptions; shares openness but applies unknown territory not multi-domain spacebreathe— micro-pause reset when collapses; single conscious breath before re-expandheal— if expansion repeats same unresolved tension → deeper self-assessment before furtheradaptic— meta-skill composing full synoptic cycle; expand-awareness = 2nd movement
GitHub リポジトリ
関連スキル
llamaguard
その他LlamaGuardは、暴力やヘイトスピーチなど6つの安全性カテゴリーにおいて、LLMの入力と出力をモデレートするMetaの70-80億パラメータモデルです。94〜95%の精度を提供し、vLLM、Hugging Face、Amazon SageMakerを使用してデプロイ可能です。このスキルを使用して、AIアプリケーションにコンテンツフィルタリングと安全策を簡単に統合できます。
cost-optimization
その他このClaudeスキルは、リソースの適正サイジング、タグ付け戦略、支出分析を通じて、開発者がクラウドコストを最適化することを支援します。AWS、Azure、GCPにわたるクラウド支出の削減とコストガバナンスの実施のためのフレームワークを提供します。インフラコストの分析、リソースの適正サイジング、または予算制約への対応が必要な際にご利用ください。
quantizing-models-bitsandbytes
その他このスキルは、bitsandbytesを使用してLLMを8ビットまたは4ビット精度に量子化し、精度の低下を最小限に抑えつつ50〜75%のメモリ削減を実現します。限られたGPUメモリでより大規模なモデルを実行したり、推論を高速化するのに理想的で、INT8、NF4、FP4などのフォーマットをサポートしています。HuggingFace Transformersと統合され、QLoRAトレーニングや8ビットオプティマイザーを可能にします。
dispatching-parallel-agents
その他このClaudeスキルは、複数のエージェントを配備し、3つ以上の独立した問題を並行して調査・修正します。共有状態や依存関係がなく解決可能な、無関係な障害が発生するシナリオ向けに設計されています。中核となる機能は並列問題解決であり、効率を最大化するために独立した問題領域ごとに1つのエージェントを割り当てます。
