Anthropic insiders say plan mode is outdated for newer Claude models. Here's what changed and the four-line brief replacing it.
What happened to Claude Code’s plan mode? #
In September, a member of the Claude Code team posted that the models probably don’t need plan mode anymore, and the idea picked up steam fast, including backing from Anthropic’s Boris on Hacker News. Plan mode had been one of the most recommended workflows in Claude Code all year, so the suggestion that it might get killed off surprised a lot of people. The real story isn’t that plan mode was a bad idea. It’s that the models changed underneath it, and the job it used to do now belongs earlier in the process, in how you write your first prompt.
TL;DR #
- Plan mode was built for older models that jumped straight into coding without thinking first, so the mode’s real function was just a prompt reminder telling the model to .
- Newer models like Opus 5.5 plan internally , deciding how much reasoning to do before acting, which makes the “don’t code yet” instruction mostly redundant.
- The value people actually got from plan mode was visibility , not the itself. Reading the plan let you catch wrong assumptions before the model burned time and tokens building the wrong thing.
- A four-line upfront brief replaces it : state what done looks like, how it will be validated, what’s out of scope, and what guardrails apply.
- Claude.txt should be audited, not deleted wholesale or endlessly appended to. Keep lines that encode decisions Claude can’t infer from your files, and cut the ones it already follows by default.
- Fixed agent teams are losing favor in both Claude Code and Codex circles, with orchestration now better handled by a single lead agent that spins up helpers only when needed.
- Cheaper models for sub-agent tasks aren’t automatically cheaper overall , since lower accuracy can mean more fix-up rounds and more total tokens and time.
Seven tools to build an app. Or just Remy. #
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Why did plan mode stop being necessary? #
Plan mode in Claude Code was never a sophisticated planning engine. It was a flag in the prompt that told the model to hold off on writing code until a plan was approved. That mattered a lot when models tended to start building immediately on a vague request, often guessing wrong about scope, files, or intent. Newer models, including Opus 5.5, behave differently: they think before responding and calibrate how much reasoning a task needs on their own. They’re also better at reading through a codebase, making reasonable assumptions, and staying on track through longer jobs without losing context. The explicit “don’t code yet” instruction doesn’t add much when the model is already pausing to reason internally.
What was plan mode actually good for, then? #
The most liked reply to the original “kill plan mode” post put it simply: plan mode wasn’t for the agent, it was for the person using it. The plan Claude produced was a readable list of assumptions, which files it intended to touch, what it considered “done,” and what it planned to change structurally. That gave users a checkpoint to catch bad assumptions before any code got written, which is far cheaper than undoing an hour of agent work after the fact.
The catch is that this workflow often papered over a weak initial prompt. A typical flow looked like: type a short request, enter plan mode, read the plan, notice the model guessed wrong about something, correct it, approve, then let it run. The brief was thin, and plan mode was doing the work of surfacing missing information after the fact. Newer models can handle the execution planning (finding the right files, templates, and order of operations) on their own. What they still can’t do is read your mind about scope and constraints that aren’t visible in your files.
What’s the four-line brief that replaces plan mode? #
Instead of relying on a plan-and-correct loop, the fix is to front-load the brief with the information a model can’t infer on its own. That means your first message should include four things alongside the actual request:
- What done looks like. Describe the end state concretely, for example matching an existing layout or including a specific new section in a specific place.
- How it will be validated. State the check that confirms success, such as totals matching a reference sheet.
- What’s out of scope. Call out anything adjacent that shouldn’t be touched, even if it seems related.
- Guardrails. Note any folders, files, or systems that are off-limits.
If you don’t know the answers yet, you can ask Claude to interview you first and extract them through questions. If you still want visibility into the model’s assumptions before it starts, you can ask for a short one-line plan rather than invoking full plan mode. And if Anthropic does deprecate the slash command, the underlying habit, writing a tight brief with explicit constraints, still works regardless of what the interface is called.
Other agents ship a demo. Remy ships an app. #
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Should you delete or keep your Claude.txt file? #
Advice on Claude.txt has flipped multiple times over the past year, from “add a line every time Claude gets something wrong” to “wipe it every six months and start over.” Both came from the same source inside Anthropic, which tells you the right answer is somewhere in between.
There’s a real case for trimming aggressively. When newer Claude models shipped, Anthropic reportedly cut a large majority of Claude Code’s own system prompt with no measurable drop in coding performance, and current guidance is to keep Claude.txt under roughly 200 lines, since longer files make the model start ignoring instructions. Separately, research on agent context files found that files generated by other tools made agents perform slightly worse while costing more to run. Every line has a token cost whether or not it’s actually helping.
But deleting everything throws out rules that were doing real work, like team-specific decisions or client preferences the model has no way of inferring from the codebase. The useful framing: deleting treats a quality problem as a quantity problem. The actual fix is going line by line and asking whether removing a given instruction would cause Claude to make a mistake. A practical starting point is running a prompt audit command to flag outdated instructions or references to files that no longer exist, then manually reviewing what’s left. If you maintain both a Claude.txt and an agents.txt, note that Claude Code will read agents.txt when no Claude.txt is present, so one file can cover both tools.
Are multi-agent teams still worth setting up? #
Opinion is split here, even among the teams building these tools. On the Codex side, the stance has been that running more than a couple of agents at once usually burns tokens without improving quality, because agents don’t trust each other’s output and end up re-checking everything, a “coordination tax.” On the Claude side, Anthropic has pushed in the opposite direction, enabling very large numbers of agents to run on a single job on its hosted platform, with internal testing showing that larger multi-agent runs caught meaningfully more planted bugs in a test codebase than a single agent did.
Where both sides agree: start small, because multi-agent setups burn tokens fast, and the model should be the one deciding how to split work, not a fixed team structure you define upfront. In practice this looks like giving a task to one orchestrating agent and letting it create sub-agents only when the job actually calls for it, rather than pre-building a “team” of a researcher, writer, and reviewer. Custom agents that are just role descriptions (“you are a senior copywriter”) can usually go. Ones that carry information Claude can’t otherwise know, like a specific checklist or a restricted toolset, are still worth keeping.
Does switching to a cheaper model for sub-agents actually save money? #
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Not always. Cheaper models look appealing for sub-agent tasks like file searching, especially given large price gaps between tiers. But cheaper per-token pricing doesn’t mean cheaper per completed task. On benchmark comparisons, a cheaper model can score meaningfully lower than a more capable mid-tier one, which means more wrong answers get passed back to the main model for correction. If a task needs two or three fix rounds because the cheap model got it wrong, you can end up spending more total tokens and more wall-clock time than if you’d used a stronger model from the start. The way to check is to run the same task both ways and compare total usage and time at the end, not the sticker price per token.
Do plugins like Superpowers still help on newer models? #
Frameworks that add structured planning, building, and testing steps on top of Claude Code became popular partly because earlier models skipped steps without that scaffolding. With newer top-tier models planning more reliably by default, the value of these add-ons is narrower than it was. Informal testing comparing several popular skill packs against a placebo file of similar length found that most didn’t outperform the placebo on Opus 5.5. That’s not necessarily a knock on any individual framework. It likely reflects that these tools helped surface exactly the gaps that model developers then fixed. The practical move isn’t to uninstall everything, but to test your actual workflows with and without a given plugin, using a clear brief either way, and keep only the ones that demonstrably earn their complexity.
Frequently Asked Questions #
Is plan mode actually being removed from Claude Code?
As of the discussion driving this debate, Anthropic has signaled it may deprecate plan mode’s prominence rather than remove the underlying command outright. The slash command is expected to remain usable even if it’s no longer the default recommended workflow.
What should I put in a Claude.txt file if I keep one?
Keep instructions that encode information Claude can’t get from your codebase, like team conventions, client preferences, or decisions already made elsewhere. Remove anything the model already does correctly without being told, and keep the file short since long files cause models to ignore parts of it.
Are agent frameworks like Superpowers obsolete now?
Not obsolete, but less necessary by default for many tasks since newer models handle more planning and structure on their own. Testing your specific workflow with and without the framework is the only reliable way to know if it still adds value for your use case.
Is it cheaper to run sub-agents on a lower-cost model?
Only if accuracy holds up. A much cheaper model can still cost more overall if it produces errors that require multiple correction rounds from the primary model, increasing both total tokens and completion time.
What’s the simplest replacement for plan mode?
Write a four-line brief with your request: what done looks like, how success will be validated, what’s explicitly out of scope, and what guardrails the model must respect. This front-loads the clarifications that plan mode used to surface after the fact.