codeclimber

I agree, that was a weak analogy. Magnus stays employed because chess fans value watching humans compete, not because engines didn't replace his capabilities.

I've updated the post title to "Train with coding assistants like Magnus Carlsen trains with chess engines" to focus on the main point: the methodology. Magnus uses chess engines as sparring-partners to improve his game after matches. Same can be done by developers who will use coding assistants to level up their skills.

Thanks for calling this out.

This is advanced/cyborg chess in a nutshell (https://en.wikipedia.org/wiki/Advanced_chess)

Companies automate the parts that are commodity. On messy product work (drifting specs, integration, liability), human + AI + good process > AI alone. The machine proposes; the human sets goals, constrains risk, writes/reads tests. That combo ships faster and with fewer costly mistakes than letting an ai free-run.

Yes, chess engines review your games and point out blunders. They also suggest moves you'd never consider. Like when you're analyzing a position and the engine recommends a move that flips the eval from -1.5 to +1.8. Similarly, coding assistants might suggest a solution you'd never considered.

Both teach something new.

Time is measurable, but they're measuring the wrong interval. Speed to first draft means nothing if every subsequent change takes longer because the code is unmaintainable. The technical debt from slop compounds: you're not saving time, you're borrowing it at a terrible interest rate.

In the long run, the LLM that produces less slop per request wins.

I've definitely thought of that clip during my own debugging sessions.

Don't run "npm run dev" during development

(I'm running it myself in a separate terminal window, otherwise there is going to be too many dev servers running simultaneously)

Well, Apple declared it an obsolete product, I doubt they will start selling them again

Thanks for the feedback! Most techniques actually work across all coding tools, though you're right I use Claude Code in examples since I'm using it mostly in real life. Which tool do you use? Happy to add more diverse examples.

I pulled ideas from posts by Simon Willison, Orta Therox, the Anthropic team, Indragie Karunaratne, Chris Dzombak, and others, then tried them on real projects with Claude Code and Codex CLI. This page keeps only what proved useful, organized by stage (planning → coding → debugging → tests → refactoring). New techniques and real‑world experiences welcome.

Creator of the website here. It’s a practical snapshot of what seemed to work with Claude Code this summer, arranged by stage and linked back to the originals. Most of it should translate across agentic tools; a handful is Claude Code–specific. Corrections and additions appreciated.

My datapoint (Claude Code): on the paid Pro tier I’d hit the cap after ~1–2 hours of real coding. I upgraded to Max and haven’t hit limits since, even with multiple terminals in parallel (lots of refactoring/tests)

Community-maintained comparison of AI coding tools with actual free access to frontier models (GPT-5, Claude Sonnet 4, Gemini 2.5 Pro, Qwen3-Coder-480B). Only includes models scoring >60% on SWE-bench Verified.

Turns out Qwen Code gives 2,000 requests/day to Qwen3-Coder-480B while Rovo Dev CLI gives 5M tokens/day of Claude Sonnet 4 during beta, among others.

The catch: every tool counts differently (requests vs tokens vs credits), and the number of requests to solve particular coding problems varies significantly from tool to tool. Makes comparison quite hard.

Quick survey: How many actual coding hours do you get from your coding tool? (Free or paid, rough numbers fine)

Nice write-up, especially the point about mixing different models for different stages of coding. I’ve been tracking which IDE/CLI tools give free or semi-free access to pro-grade LLMs (e.g., GPT-5, Claude code, Gemini 2.5 Pro) and how generous their quotas are. Ended up putting them side-by-side so it’s easier to compare hours, limits, and gotchas: https://github.com/inmve/free-ai-coding

So far I have seen GPT-5 models only in Cursor (free for paying users). Any others?

Could you share the hardware specs you use to run Devstral?

Gemini CLI (free tier) [a] - less than 1 h/day Claude Code (pro) [a] - 2 h/day Cursor (pro) [b] - 2 weeks/month

Trying to track which agentic tools (both IDEs and CLIs) give free access to pro-grade LLMs (Claude Sonnet/Opus 4, Gemini 2.5 Pro, Grok 4, OpenAI o3, etc). May be gaps or mistakes - feedback welcome.

I haven’t personally run into the fallback to Flash (probably because I mostly use Claude Code now). Curious if others using Gemini CLI are seeing the same: does the switch to Flash happen after less than an hour of coding with Gemini 2.5 Pro? Going to dig a bit and see if this is a common pattern.

Thanks, added Warp to the list with the following data:

Up to 150 AI requests/month → about 15 coding hours/month (assuming 10 req/hour) with frontier models. The Free plan explicitly supports frontier models like Claude Sonnet 4, OpenAI o3, and Gemini 2.5 Pro.

That puts Warp at #2 on the list, right below Gemini CLI.