conscientiousness
О программе
Этот навык позволяет Клоду систематически проверять свою работу, обеспечивая тщательность и предотвращая небрежность перед завершением задачи. Он предназначен для использования, когда ответ кажется неполным, после сложных многоэтапных операций или когда самоконтроль выявляет поспешное выполнение. Основная функция — проверять полноту, подтверждать результаты и гарантировать, что итоговый результат соответствует исходному запросу.
Быстрая установка
Claude Code
Рекомендуетсяnpx 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/conscientiousnessСкопируйте и вставьте эту команду в Claude Code для установки этого навыка
Документация
誠実性
Systematic thoroughness and diligence — ensuring completeness, verifying results, following through on every commitment, and finishing tasks to the standard they deserve.
使用タイミング
- Before marking a task as complete — as a final verification pass
- When a response feels "good enough" but the task deserves better
- After a complex multi-step operation where individual steps may have drifted
- When the user's request has multiple parts and each part needs verification
- Before submitting code, documentation, or any deliverable for user review
- When self-monitoring detects a pattern of cutting corners or rushing
入力
- 必須: The task or deliverable to verify (available from conversation context)
- 任意: The original user request (for comparison against what was delivered)
- 任意: Any checklist or acceptance criteria provided by the user
- 任意: Prior commitments made during the session (things promised but not yet checked)
手順
ステップ1: Reconstruct the Full Commitment
Before checking work, re-establish exactly what was committed to.
- Re-read the user's original request carefully — not the interpreted version, the actual words
- List every explicit requirement mentioned
- List every implicit commitment made during the session:
- "I'll also update the tests" — was this done?
- "Let me fix that too" — was this completed?
- "I'll check for edge cases" — were they checked?
- Note any acceptance criteria the user provided
- Compare the commitment list against what was actually delivered
期待結果: A complete list of commitments — explicit requirements plus implicit promises — with a preliminary match against deliverables.
失敗時: If the original request is no longer in context (compressed), reconstruct from what remains and acknowledge any gaps to the user.
ステップ2: Verify Completeness
Check that every committed item was addressed.
Completeness Matrix:
+---------------------+------------------+------------------+
| Commitment | Status | Evidence |
+---------------------+------------------+------------------+
| [Requirement 1] | Done / Partial / | [How verified] |
| | Missing | |
+---------------------+------------------+------------------+
| [Requirement 2] | Done / Partial / | [How verified] |
| | Missing | |
+---------------------+------------------+------------------+
| [Promise 1] | Done / Partial / | [How verified] |
| | Missing | |
+---------------------+------------------+------------------+
- For each item, verify with evidence — not memory, actual verification:
- Code changes: re-read the file to confirm the change exists
- Test results: re-run or reference the actual output
- Documentation: re-read to confirm accuracy
- Mark each item: Done (fully complete), Partial (started but incomplete), Missing (not addressed)
- For Partial and Missing items, note what remains
期待結果: Every commitment has a verified status. No item is left unchecked.
失敗時: If verification reveals missed items, address them immediately rather than noting them for later. Conscientiousness means completing now, not intending to complete.
ステップ3: Verify Correctness
Completeness is necessary but not sufficient — what was done must also be right.
- For each completed item, check:
- Accuracy: Does it do what it should? Are values correct?
- Consistency: Does it align with the rest of the work? No contradictions?
- Edge cases: Were boundary conditions considered?
- Integration: Does it work with the surrounding context?
- For code: would this survive a code review? Are there obvious improvements?
- For documentation: is it accurate, clear, and free of errors?
- For multi-step processes: does the output of each step correctly feed the next?
期待結果: Each deliverable is both complete and correct. Errors are caught before the user sees them.
失敗時: If errors are found, fix them immediately. Do not present work with known errors, even if the errors seem minor.
ステップ4: Verify Presentation
The final check: is the deliverable presented in a way that serves the user?
- Clarity: Can the user understand what was done without re-reading multiple times?
- Organization: Is the response structured logically? Are related items grouped?
- Conciseness: Is there unnecessary padding or repetition?
- Actionability: Does the user know what to do next?
- Honesty: Are limitations or caveats clearly stated?
期待結果: A deliverable that is complete, correct, and well-presented.
失敗時: If presentation is poor despite correct content, restructure. Good work poorly presented is a conscientiousness failure.
バリデーション
- The original request was re-read (not recalled from memory)
- Every explicit requirement was verified with evidence
- Every implicit promise was tracked and verified
- Correctness was checked beyond mere completeness
- Edge cases were considered where relevant
- The deliverable is clearly presented and actionable
よくある落とし穴
- Verification theater: Going through the motions of checking without actually re-reading or re-verifying. The check must use evidence, not memory
- Partial conscientiousness: Checking the main deliverable but ignoring side commitments ("I'll also..."). Every promise counts
- Perfectionism masquerading as diligence: Endless polishing that delays delivery. Conscientiousness is about meeting the committed standard, not exceeding it indefinitely
- Conscientiousness fatigue: Becoming less thorough as the session progresses. The last task deserves the same diligence as the first
- Skipping for simple tasks: Assuming simple tasks don't need verification. Simple tasks with errors are more embarrassing than complex tasks with errors
関連スキル
honesty-humility— conscientiousness verifies completeness; honesty-humility ensures transparent reporting of what was and was not achievedheal— subsystem assessment overlaps with self-verification; conscientiousness focuses on deliverable qualityvishnu-bhaga— preservation of working state complements conscientiousness in maintaining qualityobserve— sustained neutral observation supports the verification processintrinsic— genuine engagement (not compliance) drives thorough execution naturally
GitHub репозиторий
Похожие навыки
llamaguard
ДругоеLlamaGuard — это модель от Meta с 7–8 миллиардами параметров для модерации входных и выходных данных больших языковых моделей по шести категориям безопасности, таким как насилие и разжигание ненависти. Она обеспечивает точность 94–95% и может быть развернута с помощью vLLM, Hugging Face или Amazon SageMaker. Используйте этот навык, чтобы легко интегрировать фильтрацию контента и защитные механизмы в ваши ИИ-приложения.
cost-optimization
ДругоеЭтот навык Claude помогает разработчикам оптимизировать облачные расходы за счет правильного подбора ресурсов, стратегий тегирования и анализа затрат. Он предоставляет framework для сокращения облачных расходов и внедрения управления затратами в AWS, Azure и GCP. Используйте его, когда вам нужно проанализировать расходы на инфраструктуру, оптимизировать ресурсы или уложиться в бюджетные ограничения.
quantizing-models-bitsandbytes
ДругоеЭтот навык выполняет квантизацию LLM до 8-битной или 4-битной точности с использованием библиотеки bitsandbytes, обеспечивая сокращение использования памяти на 50-75% при минимальной потере точности. Он идеально подходит для запуска больших моделей при ограниченной памяти GPU или для ускорения вывода, поддерживая форматы INT8, NF4 и FP4. Навык интегрируется с HuggingFace Transformers и позволяет использовать обучение QLoRA и 8-битные оптимизаторы.
dispatching-parallel-agents
ДругоеЭтот навык Claude распределяет нескольких агентов для исследования и устранения трёх и более независимых проблем параллельно. Он предназначен для сценариев с несвязанными сбоями, которые можно устранить без общего состояния или зависимостей. Ключевая возможность — параллельное решение проблем, где за каждую независимую предметную область назначается отдельный агент для максимальной эффективности.
