The AI Coding Tools I Actually Use (And the Ones I Don't)

Cool! So, coding tools! We all use them right? I’ve been through a few of them! As a team lead working on a large SvelteKit monorepo, I’m constantly evaluating tools that could help the team move faster and write better code. Here’s my honest take on what actually works and what’s just hype.

So, if you read my posts you’ll know I post about using these tools, like, all the most popular posts on this site are now dominated by the posts I have made on coding tools! 😅

The thing is, these tools will promise a lot from the hype you see on the socials etc, but actually using them in a real production environment is a whole different story.

The Svelte 5 reality

Here’s the thing that really matters for my work: Claude 4 is really the only LLM that ‘gets’ Svelte 5 code and can write it effectively. The training data cutoff issue is real - most LLMs are still stuck on Svelte 4 patterns. Claude 4 will still default to Svelte 4 sometimes if you’re deep into a prompt and you’re not specifically specifying “Svelte 5” in your prompts.

When Windsurf lost Claude 4 access, it essentially became useless to me. Although Windsurf has a much better MCP integration than Cursor, now with Claude 4 there’s less reason for me to use MCP tools (I essentially used MCP tools to validate the Svelte 5 code being written).

Windsurf got cut off officially:

• Anthropic officially cut off Windsurf’s direct access to Claude models in June 2025
• Windsurf CEO stated: “Unfortunately, Anthropic did not provide our users direct access to Claude Sonnet 4”

I was happily using Claude 4 in Cursor for a couple of weeks after Anthropic released it. I have seen this scenario play out at least twice now, where I believe the code editor teams intercept the system prompt and add in their own crap!

Cursor’s decline was the last nail in the coffin for me. I was happy with working with Claude 4 in Cursor, but when the Cursor team flicked the “you now dumb af” switch I couldn’t use it anymore.

I’d seen stuff posted elsewhere in the past about this happening so I asked around.

My suspicions validated

I wasn’t sure if it was just me, I asked the rest of the team and they felt the same, then I had a poke around on Reddit and the Cursor forums and found that it wasn’t just me, Cursor had really gone downhill.

Cursor’s Claude 4 degradation is real:

• Users reporting Claude 4 “extremely slow since yesterday” with 0.3 tokens per second (essentially unusable)
• “Context problems got noticeably worse after the 0.45.4 update” with AI “losing context immediately”
• Multiple reports of Claude 4 working great then suddenly becoming “nearly unusable for productive work”

System prompt interference is documented:

• Users directly asking Cursor: “Did you guys change the backend system prompt?” with reports of degraded experience
• Cursor makes “frequent updates to system prompts (approximately monthly)” that disrupt user workflows
• Evidence that “System prompt makes Claude go off script and ignore rules”
• Technical analysis shows “If all models are failing in the same way, it’s likely something Cursor is injecting ahead of each prompt”

The perfect example? I asked Cursor to write tests for a file and it randomly wrote tests that didn’t even correspond to the file. When I corrected the prompt, the response was “yes I should read the file to know what I’m testing.” Unacceptable!

The Cursor team’s response? Radio silence. Despite extensive user documentation of these issues, they avoid addressing the systemic problems. When pressed with concrete evidence (users comparing Cursor vs web Claude performance), they’ll credit back requests but won’t acknowledge the underlying issues.

The wrapper problem

Here’s my theory: tools like Cursor and Windsurf are essentially wrappers around the actual LLMs. The teams behind these wrappers want to keep you engaged with their tool, so they subtly limit the effectiveness to encourage more interaction, more iterations, more usage.

Think about it - if Claude worked perfectly in Cursor on the first try every time, you’d spend less time in the tool, generate fewer API calls, and potentially use fewer credits. But if it’s “almost right” and requires tweaking, you stay engaged longer.

This is why I’m gravitating toward tools that go straight to the API:

• Claude Code: Direct Anthropic integration, no wrapper interference
• Cline: BYOK approach, your API key, your usage patterns
• GitHub Copilot with Claude 4: Microsoft’s wrapper, but they seem to be hands-off with the Claude integration

These tools have no incentive to artificially limit effectiveness because they’re not trying to maximize your session time - they just want to provide the best possible experience.

What actually works for production teams

After evaluating with the CTO, here’s what we’re actively considering:

Claude Code - The Team Choice:

• Minimal misunderstanding on tasks
• Stays focused and doesn’t get distracted by config changes (huge win!)
• Pro subscription at ~$20/month makes it cost-predictable for teams
• Direct API access means no system prompt interference
• Claude Code will stay on task most of the time and not wander off into config rabbit holes

GitHub Copilot - The Surprise Contender:

• Now has Claude 4 access in VS Code
• Appears unhindered by Microsoft’s system prompt meddling
• Seems to maintain Claude’s original capabilities
• Good fallback when you run out of Claude Code tokens

Cline - The token monster:

• “Just gets it” like Claude Code
• BYOK (bring your own key) approach via OpenRouter
• Minimal system prompt interference from the Cline team
• Can be cost-prohibitive due to token burn rate

The team impact

This isn’t just personal frustration. When tools don’t follow rules files, they write shitty tests that cause the CI to fall over. When they get distracted by config changes instead of staying on task, velocity drops. When they randomly generate code that doesn’t match the actual file structure, developers lose trust.

As a team lead, I need tools that are consistent and reliable. The current crop of “AI-enhanced” editors that meddle with system prompts are actively hurting productivity.

The bottom line

The tools that work are the ones that:

1. Go straight to the API without system prompt interference
2. Have access to Claude 4 for modern framework support
3. Stay on task without getting distracted by every config file
4. Are cost-predictable for team deployments

Right now, that’s Claude Code for team work, Cline for power users who don’t mind burning tokens, and surprisingly, GitHub Copilot as a backup option.

Everything else? Skip it. The hype isn’t worth the frustration when you’re trying to ship real code with a real team.

Your mileage may vary, but this is what works when you need to get stuff done, not just tinker with cool demos.