Learning Anti-Patterns

Avoid these patterns.

Course Collecting

Problem: Completing videos without building anything.

Fix: Every course chapter must produce a note, code sample, diagram, or project improvement.

Framework First

Problem: Using a framework before understanding the underlying concept.

Fix: Build one tiny implementation from scratch before using the framework.

No Baseline

Problem: Training complex models without a simple comparison.

Fix: Always create a naive baseline first.

No Evaluation

Problem: Judging AI output by whether it looks good.

Fix: Create golden datasets, metrics, and regression reports.

Prompt-Only Security

Problem: Depending on prompts to enforce security.

Fix: Enforce permissions in retrieval, APIs, tools, and infrastructure.

Overbuilt Architecture

Problem: Designing distributed systems before proving the core value.

Fix: Build a local vertical slice, then add infrastructure.

Ignoring Cost

Problem: Treating model calls, GPUs, storage, and logs as free.

Fix: Track cost per request and monthly cost estimates.

No Documentation

Problem: Building something that cannot be understood later.

Fix: Write README, architecture notes, evaluation reports, and ADRs.

Skipping Failure Cases

Problem: Testing only happy paths.

Fix: Add adversarial, invalid, empty, slow, and permission-denied cases.

Moving Too Fast

Problem: Starting the next topic without finishing the current artifact.

Fix: Use the module completion checklist before moving on.