I’ve been using Claude as my primary AI coding assistant for the past 4 months across Python, TypeScript, and FastAPI projects. Not ChatGPT, not Copilot inline — Claude via the web app and Claude Code CLI.
Here’s what actually changed my workflow (and a few things that still frustrate me).
1. Upload the whole file, not snippets
This sounds obvious but I kept defaulting to pasting 20-line snippets like I did with ChatGPT. Claude’s context window is massive and it performs dramatically better when it can see the full picture.
Before:
"Here's my function, why isn't it working?"
[pastes 15 lines]
After:
[uploads entire service.py + the test file]
"The test on line 47 fails. Why?"
The difference isn’t subtle. With full context, Claude catches things like an unused import that’s shadowing a variable, or a type mismatch happening 3 files away. With snippets, it guesses.
2. Ask it to plan before coding
My biggest unlock: for anything non-trivial, I ask Claude to write the plan first, not the code.
"I need to add rate limiting to my FastAPI app.
Don't write code yet. Outline the approach,
the trade-offs, and where it should live in my
existing architecture."
It’ll give you 2-3 options with trade-offs. You pick one, then ask for the code. The output is 10x better because it’s not guessing at your intent.
This also catches architectural mistakes before you’re 200 lines deep. I’ve scrapped plans at this stage probably 30% of the time — which saved me from writing code I’d have to throw away.
3. “Explain why, not just what” changes everything
Default Claude explanations are good. But adding “explain the reasoning, not just the solution” makes it a teacher.
Example: I was debugging a race condition in async Python code. Instead of just fixing it, I asked:
"Fix this, but explain why my original approach
created the race condition, and what mental model
I should use to avoid this pattern in the future."
I learned more about asyncio in that one exchange than in weeks of docs.
4. Claude Code is the underrated killer feature
If you’ve only used Claude in the browser, you’re missing the best part. Claude Code is a CLI tool that reads your whole codebase, makes changes across multiple files, runs tests, and shows you diffs before committing.
Real example from last week: I needed to migrate an API from v1 to v2 across a FastAPI project with ~40 files. I told Claude Code the migration plan and it:
- Updated route handlers across 12 files
- Adjusted the Pydantic models
- Updated the OpenAPI tags
- Modified 8 test files
- Left the legacy v1 intact with deprecation warnings
Reviewed the diff, ran tests, done in 20 minutes. Would’ve taken me 2+ hours manually.
5. It says “I don’t know” — and that’s valuable
This is the habit I didn’t expect to appreciate. When Claude isn’t sure, it says so. Compared to models that confidently hallucinate API methods that don’t exist, this is a productivity feature.
When Claude says “I’m not certain this is the right approach for your stack — can you share the config?” I know to slow down and provide context instead of trusting a confident wrong answer.
Things that still frustrate me
To keep this honest:
- Rate limits on Pro kick in during intense sessions (~50-70 Opus messages in a few hours)
- No image generation means I still need a separate tool for diagrams
- Peak hours are slow — 15-30s response times during US business hours are common
- Occasionally over-cautious on completely reasonable requests
TL;DR
Use uploads instead of snippets. Plan before coding. Ask for reasoning, not just answers. Try Claude Code if you haven’t. Trust the “I don’t know.”
If you want the full breakdown with pricing, benchmarks vs ChatGPT, and detailed pros/cons, I wrote a longer review here: Claude Review 2026
What’s been your experience? Drop your own tips below.
Top comments (0)