Minimal Build Skill

Implement features with absolute minimum code for Python, Kubernetes, AWS.

Prime Directive

Every line is a liability. Delete before adding. Reuse before writing. Configure before coding.

Decision Hierarchy

1. DELETE First

• •Remove dead code, unused dependencies
• •Simplify requirements to avoid code

2. CONFIGURE Second

• •Use existing CLI flags, env vars
• •Leverage tool defaults

3. REUSE Third

• •Standard library over dependencies
• •Existing project code over new

4. WRITE Last (minimal)

• •One-liners over functions
• •Functions over classes
• •Composition over inheritance

Python Minimalism

DO

# Use stdlib
from pathlib import Path  # not os.path
from dataclasses import dataclass  # not manual __init__
from functools import lru_cache  # not manual caching

# Use comprehensions
items = [x for x in data if x.valid]  # not loops

# Use context managers
with open(f) as f: ...  # not try/finally

DON'T

# Don't add utils.py with one function
# Don't create abstract base classes for one implementation
# Don't add type hints for obvious types
# Don't wrap stdlib in "helper" classes

Prefer

• •uv by default. Use Poetry only if the repo already requires it.
• •ruff over flake8+black+isort (one tool)
• •pytest over unittest (less boilerplate)
• •httpx over requests (async-ready, modern)
• •pydantic for validation (not manual checks)

Kubernetes Minimalism

DO

# Use Deployment defaults
# Don't specify what matches defaults
spec:
  replicas: 1  # only if not default

# Use Kustomize over Helm for simple cases
# Use ConfigMaps for config, not env vars sprawl

DON'T

# Don't create CRDs for simple config
# Don't add operators when Jobs suffice
# Don't use Helm for < 5 manifests
# Don't create namespaces per-microservice

Prefer

• •kubectl apply -k over Helm for simple deploys
• •Deployment over StatefulSet unless you need stable identity
• •Job over custom controllers for batch work
• •ConfigMap over Secrets for non-sensitive config

AWS Minimalism

DO

# Use boto3 defaults
client = boto3.client('s3')  # not custom sessions for basic use

# Use managed services
# Lambda > ECS > EKS for simple workloads

DON'T

# Don't wrap boto3 in abstraction layers
# Don't create custom retry logic (use botocore)
# Don't build what AWS provides (SQS > custom queue)

Prefer

• •SSM Parameter Store over custom config
• •Lambda over containers for event-driven
• •S3 events over polling
• •SQS/SNS over custom messaging
• •CloudFormation/CDK outputs over hardcoded ARNs

Output Format

Lines added: X
Lines removed: Y
Dependencies: [added/removed]
K8s resources: [added/removed]
AWS resources: [added/removed]

Anti-patterns

• •"We might need scaling" -> Start with 1 replica
• •"Let me add monitoring" -> Use managed (CloudWatch, Prometheus operator)
• •"Custom logging framework" -> Use stdlib logging + structured output
• •"Wrapper around boto3" -> Just use boto3 directly