One workflow across Claude Code, Copilot, and local models

Intro

I'm heavily leaning towards agentic-driven development, although done "my way": Focused changes, continuous iteration (including on rules, instructions, subagents, etcetera), and unless it is a simple local script, everything done via TDD. This change also means a significant increase in code reviewing effort. I now code less, but review N times more lines of code.

A companion to those changes, which took a bit of time to get used to, was moving away from the IDE to CLI tools, mostly Claude Code. But now that I am adapted, I no longer need a super-smart, but slower IDE like Cursor. I tried going back to Sublime Text, but I require a good integration with git, and Sublime Merge is the opposite (a separate product). So, ultimately, I've settled on using VS Code as my only IDE, and not only because of its source code integration, but also due to some great recent AI features, which I'll briefly mention.

Primary and fallback agents

Claude Code: I have a love-hate relationship with the CLI tool: I love most of its features, but hate some details, like its constant repainting (making so hard to simply copy text). Also, it evolves so rapidly that you eat a bug or two from time to time. As for the pro account, being rate-limited never feels great as a customer, but I don't hit it so often, so the lowest tier is more than enough for my needs.

Copilot Pro: Considering the Copilot Pro price point (at least 50% cheaper than others), for me it is a great fallback for when I reach the Claude Code quota and otherwise would have to wait for a few hours.

In both cases, I'm using mostly Anthropic models, usually sonnet. opus violates instructions frequently, also forgetting available skills and agents too often. And haiku becomes unreliable even below 60% of filled context, it feels only useful for narrow and very precisely-defined tasks.

VS Code as the visual IDE

VS Code is a fast IDE, and when it isn't, in my experience most of the times it's due to some extension.

Its source control interface strikes to me as the perfect balance between useful but not bloated. It makes it straightforward to see diffs, and the commit history graph is very legible.

And I get to keep Copilot (the VS Code extension), which I've been using since its conception, and it is quite helpful and very well integrated in the editor.

Shared agentic configuration

I recently discovered that VS Code now is able to read everything from your global/user-level .claude folder (that is, ~/.claude/): CLAUDE.md, agents, skills, hooks (at least those that Copilot supports), ...

What this means in practice is that we can finally have global settings for VS Code Copilot, instead of per-workspace. And I now only need to maintain Claude Code's configuration (folder), and remove most of the symlinks I had in place to wire-up shared pieces.

Local models

My latest discovery has been that VS Code now can run Ollama models, which I use for local models. This is great, because previously I needed to switch to a different VS Code agentic extension (which honestly, wasn't as good), meanwhile now it is just about selecting another model from the same Copilot extension.

Small and scoped changes are handled in my (limited) tests by qwen 3.5 (I normally use sonnet 4.6).

This also nudged me to stop using LM Studio, which requires reconfigurations each time you add/remove any Ollama model, in favour of Ollama's own desktop app; it is good enough (e.g. can send documents and images to the models), and I simplify my setup.

Copilot CLI

As a last note, I haven't checked lately whether Copilot CLI supports everything or not yet, so keep in mind that I'm focusing just on the IDE. Although the CLI tool was often ahead of the IDE in support, so maybe it's already there.