Claude reads a page in under a minute, so people assume their whole job is now fast. The reality is a thread full of retyped context, half-finished drafts and a usage limit that arrives mid-project.
The eleven habits below fix that. Each one is a setting, a project or a feature set once and used many times. Each entry names what it needs from you on the first run.
Every trick here also comes with its catch and the plan tier that unlocks it, because Claude free and Claude Pro at $20 a month teach very different lessons.
By the end you will know exactly which switches to flip today and which features to leave alone until your plan catches up.
| Trick | Where it lives | Plan or beta access | Best for |
|---|---|---|---|
| Custom instructions | Settings > General | All plans, free included | Rules that apply to every chat |
| Projects | Left sidebar | Free: 5 projects, 30MB per file; paid scales with retrieval | Background files and earlier decisions |
| Artifacts | Inside any conversation | All plans | Versioned drafts and documents |
| Effort control | Model menu in chat | Levels depend on selected model | Protecting usage on routine tasks |
| Skills | Claude settings | Introduced October | Weekly repeatable workflows |
| Claude in Chrome | Browser extension | Beta, paid subscribers | Recorded and scheduled browser tasks |
| Claude for Excel | Excel sidebar | Beta, Max, Team, Enterprise | Pivot tables and charts beside cells |
| Inline charts | Any chat response | All plans, on by default | Diagrams without exporting |
| Web search | Chat input toggle | Some level of access on most plans | Checking facts and dates |
1. Custom instructions set once
The fix costs one paste. It sits in Settings, under General, in a box labeled custom instructions. Whatever goes in that box rides along to every conversation you start, new chat or Project alike, until you change it.
Step 1 – fill the box: swap the two brackets for your own job and your own voice, then paste it.
Prompt: "You are my [role] editor. Default to [tone]. When I ask for a draft, give me a one-line summary, the draft, and three checks I should run. Ask before inventing facts, links, or quotes."
Step 2 – check it: open a fresh chat and ask for a draft with no preamble at all. If the answer arrives in your voice, with a summary above it and checks below, the setting stuck; if not, the box did not save.
Step 3 – keep it durable: the box is global, so put lasting rules there, meaning who you are, how you sound, what shape answers take, how facts get confirmed.
One-off task details belong in the chat, or they will fire on the next unrelated request. What you get back is a first answer that is usually right, from a role and tone paragraph typed once instead of nightly.
💡 Tip: Keep the saved text under 100 words. Short rules survive unrelated tasks; long ones start leaking into the wrong conversations.
⚠️ COMMON MISTAKE
Spending the window on reformatting
Running every short edit at Extra high effort is the fastest way to meet your limit on a Tuesday. Push routine formatting down to Low, and keep the top levels for the work you cannot check by eye.
2. A Project as permanent context
Never paste passwords, client files, or medical or financial records into a chat. Free and personal plans may use your chats to improve the company's models unless you switch that off in settings. Redact first, and upload only what the work needs.
Global rules fix how Claude answers, but not the files and decisions one body of work depends on.
That is what a Project holds: related chats and uploaded documents in a single place, with continuous access to those files. Claude stops meeting your background cold every morning.
Set it up in about ten minutes. Create a Project, repeat the custom instructions inside it for that subject, and add earlier artifacts so Claude can reference them in later chats in the same project.
Free is capped at five projects with 30MB per file; paid projects use retrieval to grow past that ceiling.
The payoff compounds: one workspace that answers from the same documents every time, instead of a pile of one-off chats that each start from nothing.
3. Artifact versions you can toggle
Prompt to copy
Give me two versions of this as separate artifacts. Version A: 120 words, plain and direct. Version B: 300 words with a short example. Keep every figure identical in both, and list the source for each figure underneath.
That one ask produces two editable, downloadable artifacts in the same conversation. Click between them, compare the short and the long side by side, then download the one you want.
Keep the prompt next to whichever version won, so a month later you can produce it again.
Comparing the two also exposes the slip. The two versions share figures, but a chart label or a quoted line can drift between them, or arrive invented and stated with full confidence.
Before either version ships, open every link, check each number against its source, and confirm the quote appears word for word.
The win is choice instead of archaeology. You stop scrolling three rewrites back to find the draft you liked, because both candidates sit in front of you, named and downloadable, at the cost of a few minutes of checking.
💡 PRO TIP
Ask for the version, not the rewrite
Saying 'make it shorter' buries the good draft under a rewrite. Saying 'Version B: under 90 words, same call to action' keeps both drafts alive and downloadable side by side.
4. Effort control before heavy prompts
The slow part of a long task is rarely the task. It is the heavy setting left on from last week, quietly spending usage on a reformat that a lower setting would have finished just as well.
Stored outputs save work only when the usage behind them lands on the right jobs.
The effort dial has levels Low, Medium, High, Extra high (xhigh) and Max, and the exact set depends on the model you picked.
Anthropic's guidance is plain: Low and Medium handle routine work and stretch usage, High is the best general balance, and the top levels are for complex coding and correctness-critical work.
Higher effort is not automatically better. On a simple prompt it can spend more time and more usage for no gain in the answer.
Keep the top levels for the few prompts that need them. A 700-word page still comes back in roughly 10 to 60 seconds at normal chat speed.
Hand It to Claude, or Not
- Summarizing a long PDF you already have — About a minute for the summary, then five minutes of reading it properly.
- Reformatting notes you wrote — Fast and cheap at Low effort, and you can spot an error in seconds.
- Drafting the first version of anything — Good starting material, provided you treat it as a draft.
- Client records, medical or financial files — Only with the data-improvement setting off and only what the task truly needs.
- Spreadsheet totals you will send upward — Workable, but check cell references and totals before anyone acts on them.
- A fact from memory with no source shown — Claude can state false facts with confidence and invent links. Get sources or leave it out.
- Legal or medical wording you will act on — The error is invisible to a non-expert. Use a qualified human for the final call.
- Passwords or login details — Never paste them into any chat assistant, on any plan.
5. Skills for repeatable workflows
Twelve minutes is what the weekly report costs when you rebuild the prompt each Monday.
Typed once into a skill, that prompt runs the same way every week for years.
Effort control tunes one prompt, but a workflow repeated every week needs to be stored, not re-dialled.
Skills arrived in October. They teach Claude workflows tailored to how you work, not a generic set of instructions. Claude's built-in skill creator builds one from the outcome you want.
Prompt to copy
Help me create a skill for [workflow]. Ask me for the steps, the inputs, the output format, and the checks I always make, then write the skill.
Start with one process that can be written as numbered steps, then test the finished skill on a real example before a deadline depends on it. Afterward the weekly report runs from stored instructions instead of a rebuilt prompt.
💡 Tip: Pick a process you can write as numbered steps. A vague workflow makes a vague skill.
6. Long chats without context panic
A decisions list, three open questions and one next step. That is what a three-hour thread returns when you ask for a state of the thread before continuing. Without it, the chat either forgets or starts summarizing on its own.
Anthropic changed how the context window works. Chats can now run very long, with some exceptions, by summarizing earlier messages as a thread nears its limit. A 400-line research thread that once ended mid-task and needed a manual rebuild now keeps going.
The catch is that a summary can drop detail. Restate the critical constraints yourself, and keep source files in a Project rather than trusting the thread to remember them. That reduces restarts and the copy-paste migration into a fresh chat.
By hand: about 25 minutes to rebuild the brief. The summary: about a minute. Your check: about 5 minutes.
Claude Words, Decoded
7. Recorded browser workflows in Chrome
Claude in Chrome records the browser steps you repeat each week and runs them again without you. Keeping a long chat alive only patches that conversation. This one takes the clicking itself off your hands.
The feature is in beta for paid subscribers on Pro, Team and Enterprise. Free accounts cannot use it, so check your plan before counting on it.
You perform the steps once and Claude learns them. Scheduled tasks then run the same browser jobs at the times you set.
Careful: it is beta software driving a live browser. Read every recorded action before you schedule it, and try it on a harmless page first.
A five-page Monday routine becomes a scheduled task you check in a minute. The recording takes ten minutes once, the review about five, and the Monday clicking stops.
Effort Setting: Routine Job vs Heavy Job
Keep it low
- Reformatting a table you already have
- Summarizing your own meeting notes
- Short replies and quick edits
- Routine drafting you will rewrite anyway
- Renaming and organizing files
Turn it up
- Multi-file code you will ship
- Legal or financial wording that must be exact
- Long chains of reasoning across documents
- Debugging where a wrong fix costs hours
- Anything where you cannot spot the error
8. Excel pivot tables inside Claude
Claude for Excel is beta, and it stays behind the top plans: Max, Team and Enterprise only, with pivot tables, charts and file uploads worked beside your cells.
A Pro subscriber at $20 a month cannot open it, and Max starts at $100 a month. The browser work is one place repetition hides. The spreadsheet is the other, and the bar here is higher.
For anyone already on Max, ctrl+option+c opens the full Claude app from inside Excel. The numbers and the analysis then sit on one screen, with no copied table in a second tab.
Select your data range, then paste this into the side panel.
Prompt to copy
I have selected cells B2:E40 in this sheet: monthly sales by region. Build a pivot table showing total sales per region per quarter, sorted by quarter. Add a bar chart of the four highest regions. List the cell range each total came from, and flag any row with a blank or text value where a number should be.
What comes back is a pivot table, a chart and a short list of problem rows.
Check the totals against your own sum and follow two cell ranges to their source before you trust the sheet.
The chart half of this job has a free answer, and it comes next.
9. Inline charts on every plan
A labeled bar chart of your monthly totals comes back inside the chat in under a minute, no export, no second tab, no paid tier. Claude draws custom charts and diagrams in-line by default on all plan types, free included.
Ask directly: draw a bar chart of these totals, labeled by month, and show the axis units. Paste the numbers first.
Check the axis labels, the totals and whether the chart answers the question you asked. A chart is only as accurate as the data supplied.
10. Five-hour session budgeting
Charts, browser recordings and long agent runs all draw on one usage budget, and that budget refills on a clock you can see.
Every Claude plan resets its limit on a rolling five-hour window, so the refill time moves with your first message.
Paid plans then cap you again on a weekly cycle, with bigger allowances at higher tiers.
So the limit message that killed a project at 2pm usually meant heavy runs went out at 10am, on a window already half spent.
Put the slow jobs first after a reset, since deep-research and agent runs can take half an hour. Leave formatting and short edits for later, at Low or Medium effort.
The two times: a deep-research run placed mid-window can stall for hours waiting on a reset. Placed at the start, it finishes in about 30 minutes and your edits follow in the same session.
⏱️ A session that does not burn the week
Start of window
Your first message starts the five-hour clock, so send it when you are ready to work, not to ask one question.
Heavy job first
The weekly cap sits underneath the five-hour one, so a heavy run now spends both budgets once.
Mid-window, if it failed
A limit hit after roughly three hours of use cannot refill inside this window, so switch to light edits.
Effort by job
High is the general balance; Extra high and Max are for coding and correctness work, and they draw the budget fastest.
End of window
Set a calendar reminder for five hours after that first message, so the next heavy run lands on a fresh budget.
11. Web search as a fact-check layer
Two minutes. That is how long a web search inside Claude takes to hand you a source list with dates, and it is the only step that tells you whether today's figure is real or yesterday's.
Searching keeps your answers current, not correct: every assistant, Claude included, can invent a figure, a quote or a link and say it plainly.
Step one: after any answer with numbers in it, type "Add a short source list with dates, plus a note on what you could not verify." You get a list beside the claims it backs, and a short paragraph naming the gaps.
Step two: open the links behind the three riskiest claims only, not all twenty. Check the date sits inside the window you need, and that the quote matches word for word. A search result is a lead, never a citation.
Step three: keep the source list with the work. When a client or a manager asks where a figure came from, the answer is one link with a date on it, not a re-run of the whole search.
💡 Tip: Ask 'what could you not verify' as a separate question. It gets a straighter answer than asking inside the main task.
The Eleven Tricks, by Plan Tier
🆓 Free plan covers
Custom instructions, 5 projects, artifacts, inline charts
💳 Pro unlocks
Higher usage, retrieval projects, Chrome beta
📊 Excel beta needs
Max, Team or Enterprise
⏱️ Limit reset
Rolling five-hour window
💵 Pro price
$20/month, or $17 billed annually
Start here
Custom instructions and one Project cost nothing and change every chat after them.
Frequently Asked Questions
Do I need to pay to use any of these tricks?
No. Custom instructions, five projects with 30MB per file, artifacts and inline charts all run on the free plan at $0 a month. Effort control, skills and session budgeting work too. Claude for Excel and recorded browser workflows need paid tiers.
Is it safe to upload my work files to a Project?
Only after you change the data setting. On free and personal plans, chats may be used to improve models unless that is switched off in settings. Leave out passwords, client files and medical or financial records entirely.
How long does the first setup take?
Custom instructions take about five minutes, and one Project about ten. Skills take longer, maybe half an hour for the first one, because you have to write your workflow out as numbered steps before Claude can store it.
Why do I hit my usage limit so fast?
Usually effort levels left too high on routine prompts. Low and Medium handle reformatting and summaries well and stretch usage further. Save Extra high and Max for work where a wrong answer is expensive.
Can Claude really repeat my browser clicks?
Yes, on paid plans while Claude in Chrome is in beta. It records a workflow you perform once, then repeats it, and scheduled tasks run on your timetable. Review every recorded action before you let one run unattended.
What should I set up today?
Custom instructions and one Project. Those two change every conversation that follows and cost nothing. Add effort control the first time you see a limit, then let skills and browser recording wait until a real workflow repeats weekly.
Two Settings First, Then Match the Trick to the Wall You Hit
If you retype context every morning, start with custom instructions and one Project. Those two cost nothing on the free plan and change every chat after them.
If you keep meeting your usage limit, drop effort to Low for routine work and put heavy research at the start of a five-hour window.
If your repeat work is browser clicks, charts or spreadsheets, look at Chrome, inline charts or Excel only once your plan includes them.
Set custom instructions and one Project first. Everything else waits until you know which wall you keep hitting.