Claude Code sessions: when to continue, rewind, or start fresh
Claude Code
Anthropic's AI tool for helping with coding work.
context
The chat, instructions, files, and tool results an AI can use for its next response.
context rot
Anthropic's name for possible performance decline when context becomes too large.
What drew attention
Anthropic published guidance about managing long Claude Code sessions. The subject is not simply how long an AI chat can become. It is how to decide which information should stay with a coding task.
The post describes five choices: continue a session, rewind to an earlier point, start a fresh session, compact a chat into a summary, or use subagents for separate work. The Hacker News submission received 302 points and 176 comments. Those numbers show community attention, not proof that the guidance works in every case.
The background: a session carries context
According to Anthropic's explanation, a Claude Code context includes instructions, conversation history, tool calls and results, and files the tool has read. Keeping a session can therefore help with a continuing task. The assistant retains the decisions and material already gathered.
That benefit has a limit. The company says that performance can decline slightly as context grows. It calls this possible effect “context rot”: older or irrelevant material can distract from the current task. More history is not automatically better history.
Five ways to manage the work
Continue the session when the next request is truly part of the same job. A follow-up can use files already read and choices already made.
Use /rewind when an approach failed. It returns the conversation to an earlier message. The idea is to remove the failed path from the active conversation, then give a better direction based on what was learned.
Start a fresh session when the work has become a new task. This avoids carrying unrelated details forward. There is a trade-off: for closely related work, a new session may need to read useful files again.
Use /compact to replace a long history with a summary and continue. Anthropic notes that compaction is lossy, so the summary may not preserve every detail. Subagents offer another route: give a separate part of the work its own context, then bring back its result.
Why this matters
The guidance turns prompt writing into a form of task management. People need to choose between continuity and focus. Keeping history can preserve useful evidence. Clearing it can reduce the influence of old mistakes. Summaries preserve some continuity while giving up detail.
That matters more as AI tools take on longer coding work. A user is not only asking for an answer. They are deciding what the tool should remember while it works.
What is confirmed, and what is not
The confirmed facts are limited to Anthropic's description of these session-management options and its warning about possible performance decline in very large contexts. The company presents them as practical choices, not as a universal sequence of commands.
The source does not set a single session length that is right for every project. It also does not establish that one option is best for every codebase or team. The Hacker News activity indicates interest in the topic, but it does not establish the accuracy or general success of any specific workflow.
What to watch next
The important question is whether tools can help users preserve the right details during long work. Another is how clearly an AI can show what a summary kept, what it omitted, and when human review should interrupt the task. As coding sessions become longer, managing context may become as important as writing the first prompt.
A longer AI chat is not always a better one
📰 Full story: Claude Code sessions: when to continue, rewind, or start fresh
Claude Code can use earlier work, but old details can also get in the way.
session
One ongoing conversation with Claude Code.
`/rewind`
A command that returns a session to an earlier message.
`/compact`
A command that summarizes a long session so work can continue.
💡 The gist
- Claude Code uses earlier chat and files as it works.
- Too much old information can make a task less clear.
- Users can continue, rewind, summarize, or start a new chat.
Anthropic explained how to manage long Claude Code sessions. A session is one ongoing conversation with the coding tool. The tool can use your instructions and earlier messages. It can also use files it has read and tool results.
This helps when you keep doing the same job. The tool does not need to rediscover every earlier decision. For example, it may already know the files involved.
But a very long chat can hold old mistakes. It can also hold facts that no longer matter. Anthropic says performance may decline as context grows. Context means the information the tool can see right now.
/rewind can return to an earlier message. This is useful after a failed plan. You can return before the wrong path grew larger. Then you can give the tool a better instruction.
/compact makes a shorter summary of the chat. That can make the session easier to continue. Yet a summary may leave out some details.
A new session can help with a new job. It keeps unrelated work out of the conversation. However, the tool may need to reread useful files. That is why there is no one rule for every task.
The article also mentions subagents. These are separate AI workers for separate pieces of work. They can keep one task from filling a single chat.
On Hacker News, the post received 302 points and 176 comments. That shows strong attention from that community. It does not prove every suggestion is correct.
The key question is simple: what does this task need the AI to remember?
💬 Why people disagree about Claude Code tips
People disagree about whether users or the product should do the optimizing.
- Users self-reported that reusable cache memory can unexpectedly be rebuilt, raising cost. A bug report was linked as a possible explanation.
- One user said that, in their own comparison, Codex found needed information more directly while Claude felt slower.
- Critics want the product to handle these savings automatically because its internal behavior is hard to inspect and keeps changing.
- Others say this is like renting cloud computers: learning efficient use, as with AWS or databases, is normal.
- Changing the effort setting may stop cache reuse because it changes the beginning of the prompt.
initial digest at 176 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
Give the code helper a tidy story
📰 Full story: Claude Code sessions: when to continue, rewind, or start fresh
Claude Code remembers the story of your work.
Anthropic
The company that makes Claude.
Claude Code
An AI helper for code.
Anthropic makes Claude. Claude Code helps with code.
It can remember earlier talk. It can remember files too.
That helps with one big job. But too much old talk can confuse it.
You can go back with /rewind. You can make a short story with /compact.
For a new job, start a new talk. Then old work stays away.
Hacker News noticed this post. It got 302 points and 176 comments. That means many people paid attention. It does not mean every idea is right.
💬 People want AI to be easier
They are talking about who should make an AI tool cheap and fast to use.
- Some users said the AI’s saved memory can suddenly be rebuilt and cost more. There may be a bug.
- One person felt Codex found things faster, but that is only their own experience.
- Some want the company to fix this automatically. Others say rented computers need careful use. Changing how hard the AI thinks can also stop it from reusing memory.
initial digest at 176 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
💬 Expectations and distrust around Claude Code optimization
The thread pairs frustration over users having to optimize the product with the view that operating knowledge is normal for a hosted computing tool.
initial digest at 176 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.