スキル一覧に戻る

observe-guidance

pjt222
更新日 Yesterday
3 閲覧
17
2
17
GitHubで表示
テストaidesign

について

このスキルは、行動する前にシステムを理解するための体系的な観察手法をユーザーに指導します。デバッグや研究のために、中立的なデータ収集、パターン認識、構造化された報告方法を指導します。結論を導く前に証拠を収集する必要がある場合や、証拠に基づいた分析を準備する際に活用してください。

クイックインストール

Claude Code

推奨
メイン
npx skills add pjt222/agent-almanac -a claude-code
プラグインコマンド代替
/plugin add https://github.com/pjt222/agent-almanac
Git クローン代替
git clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/observe-guidance

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

Observe (Guidance)

Coach human in field study: frame → protocol → witness → record → analyze → report. Separate fact from interpretation.

Use When

  • Person wants understand system before intervene (debug by obs, not trial-error)
  • Conducting research / evidence → needs structured method
  • Person jumps to conclusions → needs obs discipline
  • Preparing evidence-based report (not opinion)
  • Team dynamics, user behavior, process effectiveness via direct obs
  • After meditate-guidance cultivated attention → direct it at system

In

  • Required: What to observe (system, process, behavior, codebase, team, phenomenon)
  • Required: Why (debug, research, audit, curiosity, improvement)
  • Optional: Time available (single vs multi-day)
  • Optional: Prior attempts
  • Optional: Specific Qs / hypotheses
  • Optional: Recording tools (notebook, screen capture, logging, metrics)

Do

Step 1: Frame

Help set bounded frame.

  1. Ask what: "What system/behavior trying to understand?"
  2. Narrow scope: "What specific aspect interests you most?"
  3. Purpose: understanding / debug / improve / evidence / curiosity
  4. Boundaries: in/out scope → prevents endless expansion
  5. Hypothesis? state explicit, then set aside → "look for evidence both for + against"
  6. Stance:
    • Naturalist: no interfere (best for behavior)
    • Controlled: change one var, observe effect (best for debug)
    • Longitudinal: over time (best for trends)

→ Clear frame: target, scope, purpose, stance defined.

If err: can't narrow ("understand everything") → pick one entry point: "what behavior most confusing?" Already committed conclusion ("just prove X") → gently challenge: "what would disprove it?"

Step 2: Prep protocol

Systematic recording.

  1. Method by type:
    • Codebase/system: paths, line numbers, timestamps, log entries
    • Behavior/process: time-stamped notes — actor, action, context
    • Team/communication: quotes, speaker IDs, non-verbal cues
    • Natural/physical: sketches, measurements, env conditions
  2. Template:
Field Notes Template:
┌─────────────┬────────────────────────────────────────────────────────┐
│ Timestamp   │ When the observation occurred                          │
├─────────────┼────────────────────────────────────────────────────────┤
│ Observation │ What was seen/heard/measured (fact only)               │
├─────────────┼────────────────────────────────────────────────────────┤
│ Context     │ What was happening around the observation              │
├─────────────┼────────────────────────────────────────────────────────┤
│ Reaction    │ Observer's response (thoughts, emotions, surprises)    │
├─────────────┼────────────────────────────────────────────────────────┤
│ Hypothesis  │ Tentative interpretation (kept separate from fact)     │
└─────────────┴────────────────────────────────────────────────────────┘
  1. Stress separation: "obs row = fact. hypothesis row = interpretation. Never mix."
  2. Min count: "≥10 obs before any conclusion"
  3. Set up monitoring tools if applicable

→ Recording method ready. Person gets obs↔interpretation distinction. Prepared.

If err: too formal → simplify: "write what you see, separately what you think it means." Resist recording ("I'll remember") → unrecorded = memory bias; writing makes obs accurate.

Step 3: Witness

Guide actual obs session.

  1. Remind stance: "naturalist studying new species. No interfere — just watch"
  2. First 5min: pure obs no recording — just attend
  3. After immersion: begin recording w/ template
  4. Coach neutral lang: instead "system crashed" → "system stopped responding 14:32 after 47th request"
  5. Watch interpretation creeping: "that's interpretation — record in hypothesis row"
  6. Note surprises: "what surprised? surprises = most valuable data"
  7. Check frame: "still observing what set out, or drifted?"
  8. Wants to intervene: "note what + why, but don't change yet — keep observing"

→ ≥5-10 concrete obs w/ specific evidence. Experiences obs vs interpret diff. Finds harder than expected.

If err: keep interpreting → exercise: "describe as if to someone never seen this. Only verifiable facts." Run out fast → too high level → zoom in: timing, ordering, edge cases, exceptions.

Step 4: Record

Organize raw → structured.

  1. Review together
  2. Completeness: enough context for later?
  3. Factual accuracy: verifiable, or hidden assumptions?
  4. Group similar: "patterns forming?"
  5. Frequencies: how often?
  6. Absences: "what expected but not there?"
  7. Strong (clear evidence) vs weak (ambiguous)

→ Organized field notes cleanly separate obs from interpretation. Detailed enough another can verify.

If err: too vague ("things slow") → add specifics: "how slow? compared to what? which conditions?" Too detailed (record everything) → which relate to frame, which noise.

Step 5: Analyze

Obs → structured analysis.

  1. Look for patterns:
    • Repetition: "happened many times — systematic?"
    • Correlation: "X always w/ Y — related?"
    • Sequence: "A always before B — A causes B?"
    • Absence: "X never in condition Z — why?"
    • Anomaly: "all follow P except this — what diff?"
  2. Each pattern: "alternative explanation?"
  3. 2-3 hypotheses
  4. Correlation ≠ causation: "co-occur ≠ proves cause"
  5. Testable + what test confirms/refutes
  6. Confidence levels: well-supported vs speculative

→ Raw obs → structured hypotheses, data/theory separation kept. ≥1 testable hypothesis for original Q.

If err: jumps single explanation → challenge: "one possibility. another?" No patterns → too few obs → continue. Every obs same conclusion → filtering → ask: "what would contradict your theory?"

Step 6: Report

Communicate findings.

  1. Structure:
    • Context: what/when/why/conditions
    • Method: protocol, tools, duration
    • Findings: key obs w/ evidence (data, not interpretation)
    • Analysis: patterns, hypotheses, confidence
    • Recommendations: next steps (more obs, test, intervene)
    • Limitations: not covered, potential biases
  2. Findings in neutral lang separating fact from interpretation
  3. Review for hidden assumptions / unsupported claims
  4. Debug? translate hypotheses → concrete tests
  5. Report? evidence cited specifically
  6. Personal? summarize insights + remaining Qs

→ Clear report communicates obs/patterns/hypotheses, distinction kept. Reader can evaluate evidence independently.

If err: buries obs in interpretation → restructure: "facts one section, theories another." No confidence ("definitely because...") → calibrate: "how sure? what would change mind?"

Check

  • Frame set before obs (not wandering)
  • Recording protocol established + used consistently
  • Obs as facts, separate from interpretations
  • ≥5 concrete evidence-backed obs
  • Patterns from analysis, not assumed
  • Hypotheses testable, stated confidence
  • Person experienced obs-before-interpret discipline

Traps

  • Confirmation bias: only obs supporting belief. Frame must include "look for evidence against your hypothesis"
  • Intervention urge: see + fix immediately → masks root cause → observe first
  • Recording fatigue: detail = taxing. Breaks + realistic lengths (30-60min focused = substantial)
  • Over-protocol: simple obs needs notebook+timestamps. Protocol serves obs, not replaces
  • Obs ≠ surveillance: ethical boundaries matter. Visible behavior, no spy. People → transparency > secrecy
  • Skip frame: no target → attention scatters → unfocused. Rough frame > none

  • observe — AI self-directed variant
  • learn-guidance — obs feeds learning
  • listen-guidance — focused obs of speaker; obs broader to any system
  • remote-viewing-guidance — shares method adapted for non-local
  • read-garden — garden obs uses similar CRV-adapted sensory protocols

GitHub リポジトリ

pjt222/agent-almanac
パス: i18n/caveman-ultra/skills/observe-guidance
0
agentsagentskillsai-assisted-developmentclaude-codeskillsteams

関連スキル

evaluating-llms-harness

テスト

このClaudeスキルは、lm-evaluation-harnessを実行し、MMLUやGSM8Kなど60以上の標準化学術タスクでLLMをベンチマークします。開発者がモデルの品質を比較し、トレーニングの進捗を追跡し、学術的な結果を報告するために設計されています。このツールはHuggingFaceやvLLMモデルを含む様々なバックエンドをサポートしています。

スキルを見る

cloudflare-cron-triggers

テスト

このスキルは、cron式を使用してWorkersをスケジュールするためのCloudflare Cron Triggersの実装に関する包括的な知識を提供します。定期的なタスクの設定、メンテナンスジョブ、自動化されたワークフローの構築を網羅し、無効なcron式やタイムゾーン問題といった一般的な課題への対処法も含みます。開発者はこれを使用して、スケジュールされたハンドラーの設定、cronトリガーのテスト、WorkflowsやGreen Computeとの連携を構成できます。

スキルを見る

webapp-testing

テスト

このClaude Skillは、Playwrightベースのツールキットを提供し、Pythonスクリプトを通じてローカルWebアプリケーションのテストを可能にします。フロントエンドの検証、UIデバッグ、スクリーンショット撮影、ログ表示を実現し、サーバーライフサイクルを管理します。ブラウザ自動化タスクにご利用いただけますが、コンテキストの汚染を避けるため、スクリプトのソースコードを読むのではなく直接実行してください。

スキルを見る

finishing-a-development-branch

テスト

このスキルは、開発者がテストの合格を確認し、構造化された統合オプションを提示することで、完成した作業を仕上げることを支援します。実装が完了した後のマージ、PR作成、ブランチの整理といったワークフローを案内します。コードが準備できてテスト済みの際に使用し、開発プロセスを体系的に完了させましょう。

スキルを見る