conscientiousness
Acerca de
Esta habilidad verifica sistemáticamente el trabajo, comprueba su integridad y asegura que los resultados coincidan con la solicitud original antes de marcar una tarea como finalizada. Está diseñada para usarse cuando una respuesta parece "suficientemente buena", después de operaciones complejas de múltiples pasos, o para contrarrestar patrones de precipitación. Impone minuciosidad al hacer que la IA revise su propia producción frente a los compromisos, evitando así atajos.
Instalación rápida
Claude Code
Recomendadonpx 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/conscientiousnessCopia y pega este comando en Claude Code para instalar esta habilidad
Documentación
Conscientiousness
Systematic thoroughness + diligence → ensure completeness, verify results, follow through every commitment, finish tasks to standard deserved.
Use When
- Before marking task complete → final verification pass
- Res feels "good enough" but task deserves better
- Post complex multi-step op where steps may have drifted
- User req has multi parts + each needs verification
- Before submitting code, docs, or any deliverable for user review
- Self-monitoring detects pattern of cutting corners / rushing
In
- Required: Task / deliverable to verify (from conv context)
- Optional: Original user req (compare vs. what delivered)
- Optional: Any checklist / acceptance criteria from user
- Optional: Prior commitments during session (things promised but not checked)
Do
Step 1: Reconstruct Full Commitment
Before checking work → re-establish exactly what was committed.
- Re-read user's original req carefully → not interpreted version, actual words
- List every explicit req mentioned
- List every implicit commitment made during 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 from user
- Compare commitment list vs. what actually delivered
→ Full commitment list — explicit reqs + implicit promises — w/ prelim match vs. deliverables.
If err: Original req no longer in context (compressed) → reconstruct from what remains + acknowledge gaps to user.
Step 2: Verify Completeness
Check every committed item 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 | |
+---------------------+------------------+------------------+
- Each item → valid. w/ evidence, not memory, actual verification:
- Code changes: re-read file to confirm change exists
- Test results: re-run or ref actual out
- Docs: re-read to confirm accuracy
- Mark each: Done (full complete), Partial (started, incomplete), Missing (not addressed)
- Partial + Missing → note what remains
→ Every commitment has verified status. No item unchecked.
If err: Verification reveals missed items → address immediately vs. note for later. Conscientiousness = completing now, not intending to complete.
Step 3: Verify Correctness
Completeness necessary but not sufficient → what was done must also be right.
- Each completed item → check:
- Accuracy: Does it do what it should? Values correct?
- Consistency: Aligns w/ rest of work? No contradictions?
- Edge cases: Boundary conditions considered?
- Integration: Works w/ surrounding context?
- Code: would this survive code review? Obvious improvements?
- Docs: accurate, clear, free of errs?
- Multi-step processes: out of each step correctly feeds next?
→ Each deliverable complete + correct. Errs caught before user sees them.
If err: Errs found → fix immediately. Don't present work w/ known errs, even if minor.
Step 4: Verify Presentation
Final check: deliverable presented in way serving user?
- Clarity: User can understand w/o re-reading multi times?
- Organization: Res structured logically? Related items grouped?
- Conciseness: Unnecessary padding / repetition?
- Actionability: User knows what to do next?
- Honesty: Limitations / caveats clearly stated?
→ Deliverable complete, correct, well-presented.
If err: Presentation poor despite correct content → restructure. Good work poorly presented = conscientiousness failure.
Check
- Original req re-read (not recalled from memory)
- Every explicit req verified w/ evidence
- Every implicit promise tracked + verified
- Correctness checked beyond mere completeness
- Edge cases considered where relevant
- Deliverable clearly presented + actionable
Traps
- Verification theater: Going through motions of checking w/o actually re-reading / re-verifying. Check must use evidence, not memory.
- Partial conscientiousness: Checking main deliverable but ignoring side commitments ("I'll also..."). Every promise counts.
- Perfectionism masquerading as diligence: Endless polishing delays delivery. Conscientiousness = meeting committed standard, not exceeding indefinitely.
- Conscientiousness fatigue: Becoming less thorough as session progresses. Last task deserves same diligence as first.
- Skip for simple tasks: Assuming simple tasks don't need verification. Simple tasks w/ errs more embarrassing than complex w/ errs.
→
honesty-humility— conscientiousness verifies completeness; honesty-humility ensures transparent reporting of what was + wasn't achievedheal— subsystem assessment overlaps w/ self-verification; conscientiousness focuses on deliverable qualityvishnu-bhaga— preservation of working state complements conscientiousness in maintaining qualityobserve— sustained neutral observation supports verification processintrinsic— genuine engagement (not compliance) drives thorough exec naturally
Repositorio GitHub
Habilidades relacionadas
llamaguard
OtroLlamaGuard es el modelo de Meta de 7-8B parámetros para moderar las entradas y salidas de LLM en seis categorías de seguridad como violencia y discurso de odio. Ofrece una precisión del 94-95% y puede implementarse usando vLLM, Hugging Face o Amazon SageMaker. Utiliza esta skill para integrar fácilmente filtrado de contenido y barreras de seguridad en tus aplicaciones de IA.
cost-optimization
OtroEsta Skill de Claude ayuda a los desarrolladores a optimizar los costes en la nube mediante el ajuste de tamaño de recursos, estrategias de etiquetado y análisis de gastos. Proporciona un marco para reducir los gastos en la nube e implementar una gobernanza de costes en AWS, Azure y GCP. Úsala cuando necesites analizar los costes de infraestructura, ajustar el tamaño de los recursos o cumplir con restricciones presupuestarias.
quantizing-models-bitsandbytes
OtroEsta habilidad cuantiza LLMs a precisión de 8 o 4 bits utilizando bitsandbytes, logrando una reducción de memoria del 50-75% con pérdida mínima de precisión. Es ideal para ejecutar modelos más grandes en memoria GPU limitada o para acelerar la inferencia, admitiendo formatos como INT8, NF4 y FP4. La habilidad se integra con HuggingFace Transformers y permite entrenamiento QLoRA y optimizadores de 8 bits.
dispatching-parallel-agents
OtroEsta Skill de Claude despliega múltiples agentes para investigar y solucionar 3 o más problemas independientes de forma concurrente. Está diseñada para escenarios que involucran fallos no relacionados que pueden resolverse sin estado compartido o dependencias. Su capacidad principal es la resolución paralela de problemas, asignando un agente por cada dominio problemático independiente para maximizar la eficiencia.
