After 6 months of using AI coding tools daily — Claude Code, Cursor, Aider, Gemini CLI — here’s what actually makes you faster versus what wastes time.

1. Start every session with context, not code

Don’t jump straight into “write a function that…” First, tell the AI what you’re building and why:

"I'm building a SaaS billing system. We use Stripe for payments, 
Next.js App Router, Prisma + PostgreSQL. The user model is in 
src/models/user.ts. Today I'm adding subscription management."

30 seconds of context saves 10 minutes of the AI making wrong assumptions.

2. Review every change before moving on

The biggest productivity killer: letting the AI make 5 changes, then discovering the first one was wrong. Now you have to untangle 4 changes built on a broken foundation.

Review after every change. Use git diff or your IDE’s diff view. If something’s wrong, rewind immediately.

3. AI is best at the boring parts

Where AI saves the most time:

• Boilerplate (CRUD endpoints, form validation, test setup)
• Repetitive transformations (add TypeScript types to 20 functions)
• Documentation (JSDoc, README updates, API docs)
• Error handling (add try/catch, null checks, input validation)

Where AI wastes time:

• Novel architecture decisions
• Complex business logic with edge cases
• Performance optimization (it doesn’t know your data patterns)
• Debugging race conditions

4. Use two models

# Routine work: cheap/free model
aider --model ollama/qwen3:8b

# Hard problems: frontier model
claude "Debug this race condition in the payment webhook handler"

80% of coding tasks don’t need a frontier model. Save the expensive tokens for the 20% that do. See our cost management guide.

5. Don’t fight the AI — redirect it

If the AI keeps doing something wrong after 2 attempts, it’s not going to get it on attempt 3. Instead:

• Rewind to before the first attempt
• Explain what went wrong and why
• Give a more specific constraint
• Or just write that part yourself and let AI handle the rest

6. Branch before AI sessions

git checkout -b ai/add-billing
# Let AI work on this branch
# Review, squash, merge when happy

If the AI makes a mess, git checkout main and you’ve lost nothing. This removes the fear of letting AI make changes.

7. Compact proactively, not reactively

In Claude Code: run /compact when you’re 50% through the context window, not when you hit the limit. Compaction quality degrades as context fills up.

In Aider: start a new session when switching tasks. Don’t carry debugging context into a feature-building session.

8. Write the test first, let AI write the implementation

"Here's the test file (src/billing/__tests__/subscription.test.ts).
Write the implementation in src/billing/subscription.ts that makes 
all tests pass. Don't modify the tests."

This is TDD with AI. You define the behavior, AI writes the code. The tests verify correctness automatically.

9. Use AI for code review, not just code writing

After writing code (with or without AI), ask a different model to review it:

"Review src/billing/subscription.ts for:
- Security issues
- Missing error handling  
- Edge cases I might have missed
- Performance concerns"

A second opinion from AI catches things you and the first AI missed.

10. Know when to stop using AI

Some tasks are faster without AI:

• Changing a single variable name
• Adding a one-line comment
• Moving a file
• Simple CSS tweaks
• Anything that takes longer to describe than to do

The overhead of prompting, waiting, and reviewing isn’t worth it for 30-second tasks. AI pair programming is for 5-minute to 5-hour tasks.

The daily workflow that works

Morning:
  1. Open project, set context (30 sec)
  2. Plan the day's work in plain text (2 min)
  3. Start with the hardest task (use frontier model)

Midday:
  4. Switch to routine tasks (use cheap/local model)
  5. Compact or start fresh if context is getting long

End of day:
  6. AI-assisted code review of the day's changes
  7. AI-generated commit messages and PR description
  8. Write handoff notes for tomorrow's session

Total AI cost: $1-5/day. Time saved: 2-4 hours/day on a good day, 30 minutes on a bad day. The variance is high — some tasks AI nails instantly, others it struggles with for an hour.

Related: Prompt Engineering for Coding · Claude Code Cheat Sheet · Best AI Coding Tools · Tested Every Free AI Tier · AI Coding Tools Pricing · Long-Running AI Agents