Plan Mode was built for the model. The case for keeping it is about the user.
Last week, Thariq Shihipar, an Engineer working on Claude Code at Anthropic posted on X suggesting that the Claude Code Team are thinking of killing Plan Mode, a specific mode in Claude which prompts the user to fully articulate the software they want to build.
Soon influential voices in the AI Engineering space shared their views, agreeing with that direction.
I work as an AI Engineer at Indicium AI, a global, AI-native data and AI consultancy partnered with Anthropic, and I work with clients helping their engineering teams to adopt Claude Code. I’ve previously run Claude Code workshops with our clients and one of the first features we teach is how to use Shift+Tab to start Plan Mode, so naturally I reflected on these proposed changes and the reason we have Plan Mode at all.
It all starts with Peter Steinberger’s claim that Plan Mode is “for a different age”.
Why Did We Have Plan Mode?
If you’re an engineer, you’ve seen and probably experienced the magical moment when you type anything into Claude Code and the thing, which would have taken hours or days, is built. But that’s exactly the problem for Claude, you could put anything in that text field_. Anything_ from a small task to commissioning a whole new piece of software.
When Claude Code first came out, Plan Mode was a way for Claude to collect requirements for those bigger asks, and reply with questions about how you want something done. It signposted to the Agent that, this was a big task so spend time collecting the requirements, filling in the gaps in the user’s plan and breaking the requirements down into tasks. Plan Mode was specifically triggered like this so that the user could indicate the scale of the ask.
The argument against Plan Mode now, is that model reasoning has improved to the point where models can intuitively tell the difference between a big project and a small task. We don’t need to do that sign posting.
But if the Agent doesn’t need this signposting, who does need Plan Mode?
What If Plan mode is For Us?
When we think of Engineering Teams adopting Claude Code, it’s tempting to think that it’s as easy as throwing licenses and tokens at developers and letting them run wild with this powerful new technology. However, technology this capable demands new ways of working, of collaborating. It demands that we put structures in place to ensure code quality, transparency and engineer accountability.
Plan Mode provides an initial way of working for engineers on that journey and acts as a quality control measure for organisations.
Plan Mode as a Way of Working
In my experience running Claude Code Workshops and helping teams design ways of working, the art to teaching Claude Code is not just teaching the tool (documentation can do that), it’s about designing and driving the behaviours and workflows that encourage team productivity whilst also ensuring that constraints are in place so the LLM produces code of a consistently high standard.
Driving workflow adoption is hard. They usually have to be designed before they can be taught. Co-creating them with a team can help adoption, but getting leadership buy-in to do that work is often harder than getting approval for a workshop. A workshop is time-bound and built around a tangible tool. But with Plan Mode, a simple workflow is built right into the tool so it’s much easier to start that discussion.
Plan then Do. It’s a workflow that people intuitively understand and Plan Mode mirrors this, acting as a gateway to a broader discussion about ways of working.
Plan Mode as Quality Assurance
There’s an old adage in Data Science: “garbage in, garbage out”. It’s the same for Claude Code.
As discussed, this may no longer be true regarding the model’s ability, but it’s certainly true regarding the quality and detail of the request. This is the whole point of Spec-Driven Development and it’s why frameworks like /grill-me and SpecKit exist. This approach has arisen to make sure that the human drivers of Claude spend time thinking through their requirements in detail which then informs precisely how code is executed. Crucially, this also puts Engineers in a position to assume responsibility for their decisions.
This then leads to more conscious inputs to a change, better decisions and ultimately ownership over one’s work.
Yes, organisations should definitely consider using tools like Grill Me, SpecKit or OpenSpec, but Plan Mode is accessible to all from the moment that you install Claude and that sends a signal about how good software is built.
So What Should Plan Mode Become?
What if Plan Mode isn’t about putting the agent in Plan Mode? What if it is instead about putting the user in Plan Mode? What if we designed for this mantra?
As AI becomes more capable, the question shifts from “how do we get the best out of AI?” to “how does AI get the best out of us?”.
I believe it’s when we put ourselves in Plan Mode that we get the most creativity out of people. Planning, and forethought more broadly, is how we best explore and articulate a problem and dream of a solution. Planning is how we shape our future with more than just a “vibe”. So, how do improve Plan Mode to achieve this?
- Keep a Planning Tool in Claude Code dedicated to structuring the user’s planning, even if the model doesn’t need it.
- Don’t collapse planning into an effort slider. Effort is how hard the model thinks. Planning is whether the human has decided.
- Make Claude’s default behavior one of inquisitiveness that seeks to understand the user’s needs, desires and intentions.
- Where context is provided, provoke second-order questions: blast radius, rollback, who owns the decision, what “done” looks like.
- Persist decisions as a reviewable artifact for teams. This one of the benefits of /grill-me and SpecKit and it’s vital for decision traceability.
Removing Plan Mode outright risks reducing the role of human thought in defining the products shaping our future. In an age where speed is everything, having a planning tool within Claude Code makes the behaviour of pausing and planning permissible.
Alistair Heath is a Forward Deployed Engineer specialising in AI Engineering at Indicium AI. Follow me on X @alistairheath and LinkedIn. All views are my own.
Top comments (0)