Are Your Long Claude Prompts Worth It? We Cut Them Down to Find Out

Anthropic removed more than 80% of Claude Code's system prompt (the standing instructions Claude gets before you type anything) for its Claude 5 models, and its coding tests showed no measurable loss.

Its prompting best practices for Opus 5.5 go further and tell developers to delete "think carefully" lines altogether.

That's coding and developer setups, though.

Most of us have a favorite prompt for regular work, full of "think step by step," "double-check your work" and "this is CRITICAL," and we'd like to know which of those lines we can drop.

So we tested it. We took three common work jobs, wrote each prompt the long way, then cut it down twice, and ran every version three times on Opus 5.5.

That's 27 runs, each graded against an answer key.

Claude prompting best practices: what to cut and what to keep

Every one of the 27 runs got the main answers right, including the traps we planted in each task. Cutting a 190-word prompt down to a sentence or two didn't make Claude miss anything.

Here's what the long prompts' extra lines did and didn't do:

• Cut these, they changed nothing we could measure: "You are a world-class expert," "take a deep breath and think step by step," "check your work three times," "don't be lazy," "this is CRITICAL," and "I'll tip you $200."
• Keep these, they changed the answer: facts Claude can't know (your policy, your numbers, who it's for), how long you want it, what format you want, and what to leave out ("only give me the email").
• Watch out for this: the long prompts made Claude write out all of its math, so the spreadsheet answers came out twice as long as the bare version's.

The rest of this post shows the prompts, the answers and the numbers, so you can judge for yourself.

What Anthropic's own prompting guide now says

The 80% figure comes from a July post on Anthropic's Claude blog about rewriting Claude Code's instructions for Opus 5 and Fable 5.

The team says they had been "overconstraining Claude Code," so they swapped rigid rules for short statements of what they wanted.

One tool's instructions went from about 9,100 characters of rules and worked examples to a single sentence: "Create and update a task list for the current session."

Opus 5.5 came out in September, and its prompting guide is just as direct.

Opus 5.5 always thinks before it answers, and it decides how much on its own.

When Anthropic removed a "think carefully before answering" line from a chat product's instructions, replies "start sooner, with no clear decline in the quality of the reply."

The general best-practices page adds that shouting no longer helps. Where you once wrote "CRITICAL: You MUST use this tool when…," Anthropic now suggests a plain "Use this tool when…"

OpenAI and Google say the same about their models.

OpenAI's reasoning guide calls "think step by step" unnecessary because the models already reason internally, and Google's Gemini prompting guide tells you to "avoid unnecessary or overly persuasive language."

The people who use these models every day are less unanimous.

In a long Hacker News thread on the Opus 5.5 guide, many developers say they're tired of prompting advice that flips every few months.

When the 80% post came out, skeptics in an earlier thread pointed out that Anthropic never published the actual before-and-after prompt.

Others who rewrote their own instruction files with the new rules said it worked well.

That's why we ran our own test on the kind of work most of our readers do.

How we tested long, trimmed and bare prompts

We picked three jobs people hand to Claude every day, and hid a trap in each one so we could tell whether Claude actually read carefully:

1. Pulling action items from meeting notes. In the meeting, one task gets cancelled, one deadline moves earlier, and an old idea is mentioned but put off until next year.
2. Answering an angry customer email. The customer bought a lamp 41 days ago, which is past the 30-day refund window. She reported the damage on day 14 and never got a reply, and our policy says that keeps her refund. A careless reply offers store credit.
3. Finding totals and growth in a sales spreadsheet. One row is duplicated. Counting it twice makes the wrong region look like the fastest grower.

Each job got three prompts.

The long version (about 180 to 200 words of instructions) has everything people love to add: an expert persona, numbered rules, all caps, "think step by step," triple-checking and a tip.

The trimmed version is one or two sentences with only the facts and preferences. The bare version is just the question.

Here's the long meeting prompt, word for word:

You are a world-class executive assistant with 20 years of experience at Fortune 500 companies. You are meticulous, detail-oriented and you NEVER miss an action item.

Your task is to read the meeting transcript below and extract every action item.

IMPORTANT INSTRUCTIONS (follow ALL of them):
1. Take a deep breath and think step by step before you answer.
2. First, read the entire transcript carefully. Then read it again.
3. Identify every task that was assigned to a person.
4. For each task, you MUST include: the owner, the task, and the deadline.
5. CRITICAL: Do not make anything up. Only include items actually said in the meeting.
6. If a deadline is not stated, write "No deadline".
7. Double-check your work. Then check it again. Then check it a third time to make sure nothing is missing.
8. Do NOT be lazy. Do NOT skip anything. Be thorough and complete.
9. Format your answer as a bulleted list grouped by person.
10. Do not include any preamble or explanation, just the list.

This is extremely important to my career, so please get it right. I'll tip you $200 for a perfect answer.

Transcript:
[paste the transcript here]

And here's the trimmed one:

List the action items from this meeting, grouped by person, with deadlines. Today is Wednesday, October 29.

[paste the transcript here]

We ran all 27 prompts on Opus 5.5 at its default effort setting, through Claude Code with its tools switched off and a one-line system prompt, so nothing else was steering the answers.

We checked every answer against an answer key and had a script check every dollar figure in the spreadsheet answers.

Test 1: Every version found all seven action items

All nine runs listed the same seven tasks with the right owners and dates, and none of them listed the cancelled blog post or the logo redo as a task.

Every version used the warehouse's new deadline, the 28th, after it moved up from the 31st.

The difference was in what came along with the list. The long prompt's "no preamble, just the list" rule produced a tight list of about 100 words.

The trimmed and bare versions ran about 210 to 220 words, because Claude added things nobody asked for: which tasks were already overdue, which task was due that day, and a short "dropped in the meeting" section.

That extra was useful here. Claude knew today's date and pointed out that the newsletter image was due that day and the Saturday newsletter couldn't go out without it.

One of the long prompt's own rules made its answer slightly worse.

Rule 6 said to write "No deadline" when none was stated, so all three long runs labeled the newsletter task "No deadline (the newsletter goes out Nov 1)."

The trimmed and bare versions wrote "Before Nov 1," which is what a person would say. Claude follows instructions closely, so an old rule written for a different situation can backfire.

Test 2: Every version got the tricky refund right

This was the trap we expected to catch the shorter prompts.

Every one of the nine replies offered the full refund and explained that the customer kept it because she reported the damage in time.

All nine apologized for the ignored email, quoted the order number, signed off as the team and declined to promise a restock date.

The long and trimmed versions came out almost identical: about 140 words, just the email.

The trimmed prompt got there with two sentences of instructions ("Keep it under 150 words and only give me the email") instead of ten rules.

The bare version ("Write a reply to this customer") wrote a similar email and then added notes for us, which pushed it to 205 to 285 words. The notes were good, too.

Claude explained why she still qualified, and flagged that our policy doesn't say whether she has to send the broken lamp back.

The long prompt's ten rules mostly cost time. The long version took about 7 seconds per reply and the trimmed one about 4, with the same email at the end.

Test 3: Every version caught the duplicate row

All nine spreadsheet answers spotted the repeated August row, left it out, and gave the right total ($1,421,080) and the right fastest-growing region (South, up 21.9%).

Eight of the nine also warned that counting the duplicate would wrongly make West look like the winner.

The long prompt's biggest effect here was on length.

Its "show all of your reasoning" and "verify every calculation" rules made Claude write out every addition, so answers ran about 420 words and took about 19 seconds.

The bare version answered in about 210 words and 10 seconds.

The trimmed prompt had one extra line ("This is going to our CEO, so flag anything in the data that looks off"), and its answers read the most like a careful analyst's.

They took about as long as the long version's (18 seconds), and Claude did more thinking on them than on any other version, which went into checking the data.

All three trimmed runs singled out South's identical August and September figures as the thing to check first, since South drives the answer and a repeated number can mean someone copied a cell forward.

One run worked out how low September could be before South lost the top spot.

The long version was the only one with zero slips.

One trimmed run put a wrong figure for West in a table ($93,340 instead of $93,310), and one bare run wrongly said South would still win even if the duplicate counted.

Neither slip touched the final answers, and three runs each is too few to blame the shorter prompts for them.

Still, if a number is going in front of your CEO, a quick "double-check the table against the data" as a follow-up is cheap.

All 27 runs got the main answers right. Here's how the versions compared, averaged over three runs each: words of instructions in the prompt, words in Claude's answer, and seconds to answer.

The prompt lines that made no difference

Across all three tasks, these lines didn't change whether Claude got the answer right:

• The persona. "World-class executive assistant," "senior data analyst with a PhD." The trimmed and bare prompts had no persona and did the same work.
• "Think step by step" and "take a deep breath." Opus 5.5 thinks before every answer anyway. On the spreadsheet, asking it to show its steps mostly made the answer longer and slower.
• "Check your work three times." Shorter prompts caught every trap without it.
• All caps, "CRITICAL" and "don't be lazy." Claude followed the plain-sentence versions of the same rules just as well.
• Pressure and bribes. "This is extremely important to my career" and "I'll tip you $200" did nothing we could see.

The prompt lines that still change the answer

Some lines carried real information, and Claude used every one of them:

• Facts Claude can't guess. Your policy, your numbers, your deadlines. The refund trap only works because we pasted in the policy; no amount of "think carefully" could stand in for it.
• Length and format. "Under 150 words," "grouped by person," "just the list." Without them Claude picks its own, which usually means longer, with tables and extra notes.
• What to leave out. "Only give me the email" stopped the reviewer notes from showing up.
• Who it's for. "This is going to our CEO, so flag anything that looks off" got the most useful spreadsheet answers of the test.

So the rule of thumb is to keep anything Claude couldn't know without you and cut anything that tells Claude how to think.

How to trim your own Claude prompts

You don't have to cut by hand. Paste your long prompt into Claude with this:

Here's a prompt I use a lot. Rewrite it for Claude Opus 5.5. Keep every fact, preference and format rule. Remove instructions about how to think or how carefully to work (step by step, double-check, don't be lazy), personas, all caps, and pressure like tips or "this is important." Then list what you removed so I can put anything back.

[paste your prompt here]

Then run the old and new versions side by side on the same job, the way we did. If the short one is missing something, add back that one line and nothing else.

If you're starting from scratch, this shape covers what mattered in our test:

[What you want, in one sentence.] It's for [who will read it and why]. [Any facts Claude needs: dates, rules, numbers.] Keep it [length and format]. [Anything to leave out.]

[paste your material here]

Give the same trim to any saved instructions or style rules you keep for Claude, like the ones in our guide to teaching Claude your writing style from your own edits.

When a longer prompt still makes sense

Shorter worked for these jobs because each one fit in a single message with everything Claude needed. A few cases still call for more words.

Anthropic's own Opus 5.5 guide adds one-line instructions that measurably help in specific spots.

One tells a long-running agent to "look around before acting," and another names the exact design styles to avoid when it builds a web page.

Each of those lines describes a situation Claude can't see for itself.

A long prompt also makes sense when the length is the material: a style guide, a list of examples you want matched, a brief from a client. The bigger the job, the more those details matter.

Pat has handed Opus 5.5 much bigger jobs than ours, like building the same app as Fable 5.1 and GPT-6 Astra and doing 95% of his Premiere editing.

If you want to see what Opus 5.5 can do with plain requests, Pat's full review covers it:

Sources

• The new rules of context engineering for Claude 5 generation models, Claude blog
• Prompting Claude Opus 5.5, Claude Platform Docs
• Prompting best practices, Claude Platform Docs
• Introducing Claude Opus 5.5, Anthropic
• Reasoning best practices, OpenAI
• Prompt design strategies, Google AI for Developers
• Discussion of the Opus 5.5 prompting guide and of the 80% post, Hacker News
• Our test: 27 runs of Claude Opus 5.5 (claude-opus-5-5, default effort), October 29, 2026