One of the reasons I love Pi more than opinionated harnesses such as Claude Code and Opencode, is that Pi gives you unlimited flexibility. Those other tools are helpful training wheels made to implement some necessary engineering steps, using default built-in skills (see this post about creating skills: https://aibynick.com/thread/43 ). For example, opinionated harnesses tend to come with skills built in which prompt the attached model to plan out phases needed to complete bigger projects, to write tests, to choose the frameworks, tools, and libraries it thinks may end up being most effective for any piece of the project, etc., before beginning to write any code.

I tend to prefer relying on my own ability to break down a project into manageable parts, and to manage the steps intentionally, so they follow a path, and use tools/ecosystems/libraries/workflows that make sense to me. Newer, smarter models are capable of doing a lot of that planning on their own, without any rigid constraints imposed by a harness - and I don't want to be stuck with a model choosing React, for example, as my default front end framework.

In my experience, opinionated frameworks like Claude Code and OpenCode have more often ended up producing projects which weigh 50Mb to install, instead of 50Kb, and which use far more RAM to run, than I want to accept.

I'd much rather use a smart model simply to help perform research about which tools and approaches may best solve a particular problem, than rely on what's contained in the few skill files that ship with a particular harness.

That research loop is a huge part of any development process I take part in, and it's one of the most fantastic benefits of using LLMs in general to help with development work - they know so much about the vast landscape of software development, so they can help you make much better engineering choices about not only tools, but entire approaches to solving problems. That's not about writing code, but about working with greater knowledge and understanding of solutions which may be best applied to any unique situation. I don't want to outsource that endless potential capability to some limited packaged skills which come boxed in an opinionated harness.

One feature that OpenCode has built in, which may actually be helpful for groups working in a team, is automatic Git management. Aider is another Git-first coding agent which may be useful if you want to have Git versioning handled automatically (I've only just tried Aider in passing - didn't like the feel of it at all). This post may be helpful if you need to set up complex Git workflows in Pi: https://aibynick.com/thread/30

In Pi, it's easy to build your own skills and extensions to help sculp a development routine which works exactly the way you prefer, using any tools and workflows you craft, but I tend to interact with it using approaches that are devised and specifically crafted for each unique project. I don't like fighting someone else's preferred generic workflow & tooling instructions.

The other issue is that harnesses which come with large skill sets built in (such as Claude Code and OpenCode), tend to send a much larger volume of tokens to the LLM, which slows down inference and tends to pollute the context more than necessary for my needs. I'd much rather pinpoint exactly what I want the model to do, for any specific task, without the overhead of 20,000 default tokens which take the model's attention away from my specific instructions.

GPT, Claude, Deepseek, and the other frontier models have clearly been trained to think well about the engineering process, before spitting out code - and that training doesn't take up any token bandwidth - it's built into the model during its pre and post training. GPT knows how to plan on its own, and will choose tools on it's own and/or based on your instructions, or you can choose to organize a project piecemeal, as you prefer.

Just from experience, what matters most to me is the model quality and size/knowledge, and then just giving it the ability to iterate. I'd much rather have the help of an unrestricted, extremely intelligent frontier model which has endless tokens to burn on iterations (see my $20/month GPT zip file routine: https://aibynick.com/thread/29 ), than a helpful pile of skills in a harness, written by a single opinionated author, or any small group of authors. Those skills basically represent the understanding of the team who created the harness, where a model like GPT5.5 (or any other frontier model) has the combined understanding of all the humans who created all the material it's been trained on.

Start with the smartest possible model, and work with it to iterate on test results - that's what really improves a model's capability - in the exact same way that testing an application and giving humans the opportunity to iterate on those issues in a loop, improves the quality of human output.