Make Your Claude Plan Last All Month: 9 Habits That Save Usage

Know the 2 limits you're actually working against

If you've ever been mid-project when Claude told you you'd hit your limit, you know the feeling. The work stops right when it matters.

Most of the advice online about this is guesswork. So here's what Anthropic's own help center and Claude Code documentation say as of November 2026, plus what I learned from auditing several months of my own usage.

Start with how the limits work, because the fixes only make sense once you know what you're spending.

There are 2 windows. Per Anthropic's help center, Pro and Max plans have a session limit and a weekly limit. The session window starts with your first message and resets in full 5 hours later. The weekly limit resets at a fixed time each week that's assigned to your account. Max and Ultra also have a separate weekly limit for Fable, Anthropic's most capable model family, and verified Ultra accounts have no 5-hour limit at all.

It's one shared pool. Usage across claude.ai, the Claude desktop app and Claude Code all counts toward the same limits. A long chat in the browser in the morning leaves less for the Code tab in the afternoon.

How fast you spend it depends on 4 things. Anthropic lists the length and complexity of your conversations, the features you use, the model you choose and the effort level you pick. That's the important part: most of it is under your control.

You can see where you stand any time under Settings, then Usage, in Claude. It shows how much of the current session you've used, the time left in it, and when your weekly limits reset. In the Code tab, click the usage ring in the row under the prompt box to see the same plan bars.

Find out what's eating your usage first

Before you change any habits, look at the evidence. Claude Code tracks where your usage goes, which beats guessing.

On Pro, Max and Ultra plans, Claude Code's /usage command breaks down what counts against your limits. Per Anthropic's documentation, as of November 2026 that breakdown includes:

• Attribution: the share of recent usage that went to skills, helper agents, plugins and individual connected tools.
• Recommendations: one suggested change per row, each covering at least 10% of your recent usage, like trimming a long CLAUDE.md file. Long conversations and cache misses are among the patterns it flags.
• Loops: a row for each of the heaviest recurring tasks that ran inside a session, with how often it fires and how much it used.

It covers the last 24 hours or the last 7 days. The figures are approximate and come only from session history on that computer, so usage on other devices or on claude.ai isn't included.

The easiest way to act on it from the desktop app is /doctor, Anthropic's built-in checkup. Type it in the prompt box. On those plans, its usage check covers the changes /usage recommends over the last 7 days. Claude reports what it found first, and it asks before it changes anything.

Look for the biggest single line, and fix that first. One habit usually accounts for more than all the others combined.

Start a fresh chat for each new job

This is the habit with the biggest payoff for most people, and it's free.

Here's why. Claude doesn't remember a conversation the way you do. Every time you send a message, the whole conversation so far goes with it. Anthropic's documentation says it plainly: Claude Code sends your full conversation with every request.

So a chat that started with your email newsletter, wandered into a website fix, then a client proposal, then a social post, is carrying all of that history on every single message. A one-line question at the end of a long, mixed chat can cost as much as the whole morning's work.

My own audit showed how lopsided this gets. Across 5 months, the 3 biggest chats used 22% of all my usage, and the top 10% of my hands-on chats used 67% of the hands-on total. A few giant, wandering chats cost more than hundreds of small ones.

One job, one chat. When you switch to unrelated work, start a new session. Typing /clear in the prompt box does the same thing inside the current one, and per Anthropic it costs nothing. If you want to come back later, rename the chat first so you can find it again.

There's a middle option too. /compact summarises the conversation so far into a short recap and continues from there. It's useful when you want to keep going on the same job with less baggage. But compacting reads the whole conversation to summarise it, so it's a big request in itself. When you want a clean start, /clear is the cheaper move.

Per Anthropic's documentation, Claude Code sends your full conversation with every request. That's why one job per chat saves so much.

Wrap up big chats before you walk away

This one surprised me most, and it's invisible until you measure it.

When you send a message, Claude keeps a copy of the conversation ready for a while, so the next message can reuse it cheaply. That's called the cache. Per Anthropic's documentation, on a subscription within your plan's included usage, Claude Code keeps the main conversation's cache for 1 hour. Once you're running on usage credits, it drops to 5 minutes.

Come back after that, and your first message has to reprocess the entire conversation from scratch. On a short chat, that doesn't matter. On a big one, it adds up fast.

When I audited my own prompts across 5 months, I found I'd come back to a cold chat 876 times. Each return re-read about 468,000 tokens of history. Together, that was about 10% of my hands-on usage, spent purely on re-reading conversations I'd walked away from.

3 fixes, in order of effort:

1. Finish the thought before a break. If you're about to step away from a big chat, wrap up the current step first.
2. Start fresh with a handoff. Ask Claude for a short note on where things stand, save it, and start a new chat from the note tomorrow instead of reopening the giant one.
3. Let Claude Code shrink it for you. As of November 2026, the Code tab compacts a big conversation (200,000 tokens of context by default) after it has sat idle for about an hour, before the cache expires. Your next message then sends a summary instead of the whole chat. It only runs while your computer stays awake.

A summary can lose details from early in the chat, so for work you care about, the handoff note is the safer habit. Here's a handoff prompt that works at the end of any big session:

Before I close this chat, write a short handoff note for a fresh session:
what we were doing, what's finished, what's next, the files involved,
and any decisions or rules I gave you along the way.
Save it as handoff.md in this project folder. Keep it under 300 words.

Tomorrow, start a new chat in the same folder and say "Read handoff.md and continue." You get the context back for a fraction of the cost of reopening the old chat.

I wrote up the full audit, and the small add-ons I built from it, in find where your AI hours really go.

In my own audit, returning to cold chats cost about 10% of my hands-on usage: 876 returns, each re-reading about 468,000 tokens.

Spot the 4 signs a chat has grown too heavy

You don't need to check numbers constantly. A heavy chat usually tells you it's heavy, if you know what to look for.

1. Claude starts forgetting what you said early on. As a conversation fills up, Claude Code summarises it, and Anthropic warns that detailed instructions from early in the conversation may be lost. If Claude stops following a rule you set at the beginning, the chat has outgrown itself.

2. The context gauge is running high. In the Code tab, the Context card in the session details panel shows how full this chat's context window is, as a percentage. That gauge measures this one chat, separately from your plan limits, and a high number means every message is carrying a lot.

3. You're scrolling to find things. If you can't remember what's in the chat, Claude is still carrying all of it, on every message.

4. The chat covers more than one job. The moment a session has drifted from its first task to a second, unrelated one, you're paying for the first job on every message of the second.

When you see any of these, ask Claude for a short handoff note, save it in the project folder, and start a new chat from it. You lose nothing important, and every message after that is cheaper.

Match the model to the job

Not every job needs the most powerful model. Using the biggest one for everything is one of the fastest ways to drain an allowance.

Anthropic's documentation puts it simply: Sonnet handles most coding tasks well and costs less than Opus, so reserve Opus for complex decisions and multi-step reasoning. For quick, simple helper tasks, Haiku is lighter still.

Above Opus sit the Fable models, which Anthropic calls the most capable models in Claude Code, suited to tasks larger than a single sitting. Depending on your plan, Fable can bill to usage credits instead of your included usage, so check how it bills before you make it your default.

A practical way to think about it, with the current versions as of November 2026:

Pick the model in the model picker, in the row under the prompt box, before your first message. You can switch mid-chat, but each model keeps its own cache, so the first message after a switch re-reads the whole conversation. Switch at a natural break, or start a new chat with the new model.

Per Anthropic's docs, each model has its own cache: switch models mid-chat and the next request reads the entire conversation with no cache hits.

There's a second dial too, called effort. Higher effort means more thinking before each answer, which is great for hard problems and wasteful for easy ones.

As of November 2026, the current models start at medium, and you can change the level under Effort in the model picker. Anthropic says the top setting, max, "may show diminishing returns and is prone to overthinking," so raise effort only for the jobs that need it. On the 5.5 models and Fable 5.1 or later, changing effort keeps the cache.

Keep your standing instructions short

Many people keep a file of standing instructions that Claude reads at the start of every session. In Claude Code, that's the CLAUDE.md file in your project folder.

It's useful. It's also read at the start of every single session, whether today's work needs it or not. Anthropic's documentation recommends keeping it under 200 lines. Longer files, it says, "consume more context and reduce adherence."

The fix is to move procedures out of that file and into skills. A skill's full instructions load only when the job needs them.

So "how I write a client report," "how I build a carousel" and "how I check a sales page" can each be a skill, and they cost almost nothing on the days you don't use them.

Each installed skill still adds its 1 or 2 sentence description to every message, so switch off the skills you never use. If skills are new to you, start with what a Claude skill is.

A good rule from Anthropic's guidance: add a line to your standing instructions when Claude makes the same mistake a second time. Everything else belongs in a skill, a reference file, or nowhere.

Point Claude at files instead of pasting them

When you paste a long document into a chat, it stays in that conversation, and travels with every message that follows. A 20-page report pasted at the start of a session gets carried through every question you ask afterwards.

In Claude Code, put the document in your project folder and refer to it instead. Claude can then read the part it needs, when it needs it.

If you work in regular Claude chats, a project does a similar job. Anthropic's help center says content in projects is cached and counts less against your limits when you reuse it. Put the documents you keep coming back to in a project instead of uploading them to each new chat.

The same goes for vague requests. Anthropic's documentation notes that a broad ask like "improve this codebase" triggers broad scanning, while a specific one lets Claude work with minimal file reads. "Tighten the 3 headlines on the sales page in the offer folder" costs far less than "make my website better."

The help center adds another tip: batch related questions into one message instead of sending them one at a time.

Plan before big jobs

The most expensive thing you can do with AI is the wrong big thing. A large change that heads in the wrong direction costs the usage to make it, then the usage to undo it, then the usage to redo it.

Claude Code has a Plan mode for exactly this. You choose it from the mode selector in the row under the prompt box. In Plan mode, Claude reads and explores, then proposes an approach, and edits nothing until you approve it.

Reading a plan takes a minute. Catching a misunderstanding in a plan is nearly free. Catching it after Claude has rewritten 15 files costs the usage 3 times over.

And if a job does start heading the wrong way, click the stop button early. Every extra step in the wrong direction is usage you'll spend again.

Watch the work that runs while you're away

Here's a source of usage people forget about entirely: work that runs without you.

Per Anthropic's documentation, a loop that repeats inside a session fires on its interval even while the session sits idle, and sends your full context each time. Helper agents, and every agent a multi-step workflow starts, send their own requests on top of the main conversation. Even idle time can use a little, through session recaps and idle compaction.

None of these are bad. I run plenty of scheduled jobs, and they save me hours. But they draw from the same pool as everything else, so they deserve the same scrutiny:

• Put recurring jobs on the Scheduled page. In the Code tab, click Scheduled in the sidebar (it may sit under More). A local scheduled task starts a new session each time it runs, so it never drags a giant chat along with it.
• Check how often they really need to run. A report that refreshes hourly but gets read once a day is 23 wasted runs.
• Check the breakdown once a week, with /doctor or the Loops rows in /usage, and trim the heaviest item.

Per Anthropic, a scheduled task inside a session fires on its interval even while the session is idle, sending your full context each time.

Decide what to do when you still hit the limit

Even with good habits, a big week can run you into the limit. You have 3 options, per Anthropic's help center as of November 2026:

1. Wait for the reset. The session window resets 5 hours after it started, and Settings, then Usage, shows exactly when. Weekly limits reset at your account's fixed weekly time.
2. Move up a plan. Max includes 5 times Pro's usage every 5 hours and every week, and Anthropic points Max users who hit limits consistently to Ultra.
3. Turn on usage credits. In Settings, then Usage, the Usage credits section lets you keep working past a limit at standard API rates, with a monthly spend limit you set yourself.

There's also a small cushion built in. Per the help center, if you hit the 5-hour limit while Claude Code is partway through a response, it may keep working briefly to finish that step. That allowance is capped, counts toward your weekly limit, and is still rolling out as of November 2026.

A good rule: if you hit a limit once, check the breakdown before you do anything else, because it's usually one fixable habit. If I hit it every week after fixing the habits, the work has outgrown the plan, and that's a good problem.

Put the habits on autopilot

You don't have to remember all of this. The best version of these habits runs without you thinking about it.

Start a new chat for each job by default, and keep a short "where this stands" note in each project folder so fresh chats start fast. Move every repeated procedure into a skill. Run /doctor once a week, at the same time, and fix the single biggest line.

Do those 3 things for a month and you'll know exactly where your allowance goes, which is more than most people ever learn.

See exactly where your usage goes with AI Rewind

AI Rewind, inside the Solo Creators AI Studio, reads your own ChatGPT, Claude Code and Codex history on your computer and shows you how your work actually breaks down: which projects took the most, when you work, and what you keep coming back to. For Claude Code and Codex, it also totals what that work would cost at list API prices.

It's the member-ready version of the audit I ran on my own prompts. Every number opens to the real messages behind it, so you can see exactly what to change. You install it by dragging it into Claude or Codex and asking, and it comes with a short course on how to use it.