10 habits to run Claude all day on $20 plan

Last week I told you to keep Kimi away from anything client-facing, and I promised you the other half of that answer: how to get more out of a model you actually trust.

I spent this week measuring where my own Claude usage actually goes. 98% of it went on re-reading things I'd already sent. 2% produced actual work.

A PDF page costs 1,500 to 3,000 tokens. So a 15-page report is roughly 38,000 tokens before you've asked it a single question.

Paste those same 15 pages into a Google Doc, download it as markdown, upload that instead: about 2,000 tokens.

Same document. Same answer. 36,000 tokens you never had to spend.

That's the first of 10. Together they're the difference between running Claude all day on the $20 plan and hitting a wall at 3pm on the $100 one.

Most business owners are wasting tokens & hitting limits

• They blame the plan and upgrade to Max, which buys more room for the same waste (like a hoarder buying a bigger house!)
• They correct Claude in a new message, and every correction makes it re-read the whole chat before it answers.
• They upload the same PDF into 5 different chats and pay full price 5 times.
• They keep one chat open all day, so a recipe, a client proposal and next week's ad copy all share the same history.
• They run Fable on a grammar check, then wonder where the afternoon went.

How Claude actually spends your limit

A token is roughly a word.

Every time you hit send, Claude re-reads the whole conversation from the top before it answers.

Every message you wrote. Every answer it gave. All of it.

So message 1 costs almost nothing. Message 30 means re-reading 29 exchanges before it even starts thinking about the question you just asked.

The chat gets longer, and every message gets more expensive than the one before it.

Every habit below does the same job. Give Claude less to re-read.

10 habits to run Claude all day

1. Convert files before you upload (my favorite)

Paste the PDF text into a Google Doc, then File > Download > Markdown (.md). Upload the .md in Claude.

DOCX and PPTX files carry metadata bloat you can't even see. Same fix. A full-size screenshot runs about 1,300 tokens, so crop it tight and it drops under 100. And if the useful part is one number, type the number.

2. New topic, new chat

Every reply re-reads the entire conversation from the top. A 20-message session runs around 105,000 tokens. Push it to 30 messages and it's 232,000.

That's 127,000 extra tokens for 10 more messages, and most of it went on reading the first 20 again.

Long chats get more expensive and less sharp at the same time. Somewhere around 30 to 50 back-and-forths it starts contradicting itself. Past 100, just leave.

At 15 to 20 messages, ask Claude for a summary of where you've got to, paste that into a fresh chat, and close the old one. You keep the conclusions and drop the history you're paying to re-read.

Keep topics apart, too. A client proposal, next week's ad copy and a dinner recipe in one chat means every one of them is paying to carry the other two.

3. Edit your message instead of sending a follow-up

Hover over the message you sent, hit edit, fix the wording there. Claude throws away its old answer and starts again from your new version.

Send a correction instead and the bad prompt + the bad answer both stay in the chat, and you pay to re-read them for the rest of the conversation.

1 clean prompt > 5 corrections.

4. Stop asking Claude to redo the whole thing

"Only redo section 2. Keep everything else exactly as it is. No commentary, no explanations, just the output."

That last sentence matters. Without it you get 3 paragraphs explaining what it just did, and you pay for every one.

Regenerating a 2,000-word document to fix 4 sentences is the most common way business owners burn a session before lunch.

Don’t be lazy.

5. Put recurring files in a Project

Upload a file to a Project once and every chat in that Project can use it. Better still, Claude pulls out only the relevant chunks instead of reading the whole thing top to bottom.

Brand guide, pricing sheet, ICP doc, your last 3 sales call transcripts.

Then stop. Load 40 files in and it starts remixing your own documents back at you instead of thinking. Cap it around 6 to 8.

When it's your process that repeats and not your files, build a skill instead. Proposals, ad copy, weekly reports, onboarding, sales follow-ups. A skill carries the steps and only loads when you call it.

6. Ask for everything in one message

3 separate prompts mean it re-reads the whole chat 3 times. Send the whole draft for editing, not paragraph by paragraph.

Keep the brief itself short. A 500-word prompt gets re-read on every message that follows it. A 30-word prompt does the same job for a fraction of the price.

So when you're not sure what you want yet, make it ask you instead of guessing:

"I want to [task] so that [what good looks like]. Ask me questions before you start."

That one line kills the 4 rounds of "make it shorter" that follow a vague brief.

Same when you're editing. "Make it punchier" does almost nothing. "Never write like this: [paste the thing you hate]" draws the line once and saves you the next 3 rounds.

7. Switch off what you're not using

Connectors, web search, extended thinking, research mode: if they're switched on for your account, they cost you on every message whether the job needs them or not.

Keep everything off by default. Turn on what the job needs, then turn it back off.

One thing worth setting up once: Personal Preferences and Styles, in Settings. Write your tone and format rules there and you stop re-explaining them at the top of every chat.

8. Match the model to the job

Haiku for quick lookups, reformatting and yes/no questions.

Sonnet for the bulk of your writing, editing and analysis. Most of your day should live here.

Opus for strategy work and messy financials, where being wrong costs you real money. If you could check the answer yourself in under 30 seconds, it didn't need Opus.

If you're on the Max plan, run Fable as the mastermind and nothing else. Give it the messy problem, make it plan the job, break it into pieces and write the brief. Then hand the execution to Opus /Sonnet in a separate chat.

You pay Fable prices for the 5 messages that decide things, and Sonnet prices for the 50 that do the work.

You can do the same inside one chat. Frame the problem on the big model, then switch models from the dropdown and carry on. Your first prompt is the one that matters most, so that's where the expensive model earns its keep.

9. Plan in Chat, build the file in Cowork

Making an actual spreadsheet, doc or deck costs far more than talking about one.

Settle the structure, the assumptions and the numbers in a normal chat. Move to Cowork once there's nothing left to decide.

And when you get there, point it at the 2 folders it actually needs. Hand it 50 files "just in case" and it reads 50 files.

10. Spread your sessions across the day

As of August 2026, paid Claude plans run a 5-hour session limit plus a weekly limit, and  Fable (for MAX users) gets its own separate weekly line. Both are visible in Settings > Usage.

Burn the whole session before lunch and your afternoon is dead until it rolls over.

Heavy work in one session, admin in the next, and keep the third in reserve for the fire that always shows up at 4pm.

If you want help setting this up across your team, email me.

Bonus: the handoff skill by Matt Pocock

Type /handoff and tell it what the next session is for:

"/handoff finishing the Q4 pricing page"

It writes a handoff document summarising the current conversation so a new chat can continue the work. It links out to your existing files instead of copying them in, and strips out anything sensitive.

Install with npx skills@latest add mattpocock/skills --skill=handoff

Your $20 plan is bigger than you think

Most of your limit goes on re-reading what you already sent.

Do these 4 and $20 gets you further than most people get on $100:

Convert your files to .md
Start a new chat when the topic changes
Edit your message instead of sending a follow-up
Switch off connectors you're not using

If someone on your team complains about limits every week, forward them this. It's cheaper than buying them a bigger seat.

P.S. run habit 1 this week, then reply with what your usage bar looks like after 3 days. I read everything.