I've Revised My Prompts Using These Suggestions, But Claude Still Gives Me Unwanted Answers. What Should I Do?
Diagnose where the problem is, then address it specifically.
Situation 1: Claude's response direction is completely wrong — your prompt's description of the task still isn't clear enough. Solution: within the same conversation, directly say 'your response isn't what I need, let me re-explain my requirements: [re-explain].' This is more effective than starting a new conversation because Claude already has some background.
Situation 2: Claude's response direction is right, but the details are wrong (tone, length, format) — the easiest situation to fix. Say directly within the same conversation: 'the overall direction is right, but a few things need adjustment: (1) compress the second paragraph to 2 sentences; (2) shift the tone from formal to conversational; (3) change the closing to a question.' Claude can usually execute these specific revision instructions quite precisely.
Situation 3: Claude's response looks reasonable but you can't articulate why it's wrong — this usually means your own goal isn't clear enough yet. The most effective approach is: tell Claude 'what makes me unsatisfied with this answer is [describe it] — can you try again from a different angle?' Let Claude help you clarify your own requirements.
Does Role Setting (Role Prompting) Actually Work, or Does It Just Look Meaningful?
It works — and the effect is quite noticeable in specific scenarios — but not every situation requires it.
Scenarios where role setting is most effective: (1) you need specialized industry knowledge and terminology — telling Claude 'you are a senior sales executive with extensive SaaS sales experience' makes responses more grounded in that industry's real context; (2) you need a specific critical perspective — 'you are a strict investor whose job is to find weaknesses in this business plan' makes Claude more proactively problem-seeking than default mode; (3) you need a specific communication style — 'you are a teacher skilled at explaining complex concepts clearly, your students are business professionals with no technical background' makes explanations more appropriate for that audience.
Scenarios where role setting makes little difference: basic information lookup queries (e.g., 'how do you say this word in English'), or very straightforward tasks (e.g., 'translate the following Chinese text into English') — adding role setting in these cases makes minimal difference.
Conclusion: not every prompt needs role setting, but for tasks requiring professional perspective or a specific communication style, role setting is one of the lowest-cost ways to improve output quality.
I Tried Having Claude Revise Its Output, But the Revised Version Feels Worse. What Happened?
This happens occasionally, usually for two reasons:
Reason 1: Your revision instructions weren't clear enough. 'Make this better' is an unexecutable instruction — what does 'better' mean? If you can't articulate specifically what you're dissatisfied with, Claude can't make the revision you want. Solution: be more specific about what you're unhappy with: 'I think the first sentence is too formal and sounds unnatural — can you try a more conversational phrasing that preserves the same meaning?'
Reason 2: Your revision instructions conflict with the original requirements. If your original prompt said 'formal tone' and then you say 'make the tone more relaxed,' Claude may find a compromise between the two requirements, resulting in something that feels like neither. Solution: if you notice the conversation is getting confused, the best approach is to start a new conversation, consolidate all requirements into a clearer single prompt, and start fresh.
A preventive habit: before asking Claude to revise, mentally confirm 'what specific change do I want?' If you can't articulate it, Claude can't either. Having Claude revise is most effective when you already know where the problem is.
Writing Detailed Prompts Every Time Is Exhausting. Is There a Way to Avoid Re-Explaining So Much Each Time?
Yes — and this problem is one of the reasons Claude Projects and Custom Instructions were designed.
Solution 1: Claude Projects Custom Instructions. Write the background information you most often need to specify (who you are, output preferences, tone requirements) into your Claude Project's Custom Instructions — once. After that, every conversation in this Project automatically has Claude knowing these settings; you don't need to repeat them in every prompt.
Solution 2: Build a personal prompt template library. If you have a few common task types (e.g., 'weekly report organization,' 'monthly client communication drafts'), build a fixed prompt template for each task type, stored in a text document or Notion. Each time you use it, only replace the core content — the format and background specifications don't need to be rewritten.
Solution 3: Leverage conversation continuity. Within the same conversation, background you establish in the first message is retained by Claude for all subsequent messages. For complex multi-step tasks, you don't need to re-explain background for each step — stating it clearly once at the start of the conversation is sufficient.
Long-term, the most time-saving investment is configuring Custom Instructions well — an upfront 30-minute investment saves 2–5 minutes of repeated explanation per conversation thereafter.
Many people who have been using Claude for a while experience a strange phenomenon: sometimes the answers are excellent, sometimes completely off, but you can't articulate what's different. Your prompts seem roughly the same, but outputs vary wildly.
This usually isn't Claude being inconsistent — your prompts have some hidden problems. Here are the five most common prompt mistakes beginners make, along with improvements for each. Understanding these five will immediately elevate your Claude output quality.
A prompt's essence is you telling Claude: 'I need you to do this thing, in this context, in this way, to produce this result.' But most people's prompts only include one or two of these elements, omitting the rest.
When you omit 'context,' Claude assumes the most generic context. When you omit 'method,' Claude uses whatever output approach seems most reasonable to it. When you omit 'result specifications,' Claude outputs in its default format.
These assumptions sometimes align with your expectations, sometimes not at all — which is why your output quality seems unstable, when the real problem is prompts missing key information.
Mistake 1: Prompt too short, insufficient background information
The most common beginner mistake: only saying what you want, without specifying who you are, what your goal is, or where the output will be used.
Bad example: 'Help me write an email to a client.'
Problem: Claude doesn't know your relationship with this client, the email's purpose, the desired tone, or your industry background. It can only respond with the most generic business email format — usually too formulaic and hollow.
Improvement: 'I'm a design consultant. My client is the marketing director of a retail company. I submitted a brand strategy report last week and now need to write a follow-up email asking whether we can schedule a 30-minute discussion next week to confirm their thoughts on the report. Tone should be professional but warm, not overly formal. Under 150 words.'
Mistake 2: Prompt too long, stuffed with contradictory or unimportant requirements
The opposite of Mistake 1: out of fear that Claude will misunderstand, cramming every possible requirement into the prompt, leaving Claude unsure which is actually the key priority.
Common symptoms: your prompt exceeds 500 words with many 'sometimes this, but if not then that' conditional specifications; Claude's output is always slightly off from what you want, but you can't articulate why.
Improvement: identify the 3–5 most important requirements clearly and remove all 'possible exceptions' and 'less important preferences.' A good prompt is explicit, not 'covering all scenarios.'
Mistake 3: Not telling Claude what role to play
Claude defaults to being a 'general assistant.' If you need it to respond from a specific professional perspective but don't say so, it responds generally — even if that's not the best angle.
Improvement: at the start of your prompt, specify the role you want Claude to play. One sentence is enough: 'You are a marketing strategist with 10 years of retail industry experience,' or 'You are a strict editor whose job is to find weaknesses in arguments, not praise the writing.'
Role setting is most effective for: analysis requiring specific industry knowledge, evaluation requiring a critical perspective, text generation requiring a particular communication style.
Mistake 4: Not specifying the output format
'Help me analyze the pros and cons of this situation' — Claude might use bullet points, might use paragraphs, might use a table, entirely depending on its own judgment. If you have specific format requirements, you must state them explicitly.
Improvement: add format specifications to your prompt. For example: 'Please output in this format: (1) Executive summary (3 sentences max); (2) Main advantages (bullet points, max 5); (3) Main risks (bullet points, max 3); (4) My recommendation (1 paragraph).'
Common format specifications: word count range, bullets vs. paragraphs, whether headers are needed, whether examples are needed, output structure (e.g., STAR format, Problem-Cause-Solution format).
Many people use Claude in a 'one question, one answer' mode — ask a question, receive an answer, if unsatisfied start a new conversation and ask again. This is the least efficient approach.
Within the same conversation, Claude retains memory of all previous context. If its first response doesn't fully meet your needs, simply say within the same conversation: 'This answer is good, but I think the second paragraph is too long — can you compress it to 3 sentences? Also change the closing to a question that prompts readers to think.' Claude will revise directly based on your feedback, without you needing to re-explain all the background information.
Treat Claude's first response as a 'draft,' not a 'final version.' A good prompting workflow is: the first prompt establishes the foundation, then one or two rounds of revision dialogue bring the output progressively closer to what you want. This is far more efficient than re-asking from scratch, and the final result is typically better.
You don't need to fix all five mistakes at once — start with the easiest:
Change you can make today (5 minutes): before asking Claude anything next time, spend 30 seconds asking yourself 'did I clearly specify my background, purpose, and expected format?' This habit immediately improves most prompts.
Change for this week (30 minutes): in your main Claude Project's Custom Instructions, add the role settings and output format preferences you use most often. Set once, automatically applied to all subsequent conversations.
Long-term habit to build: treat everything Claude generates as a 'draft,' not a 'finished product.' Develop the habit of using one or two rounds of revision dialogue to adjust drafts to your desired version — this habit is the key to making Claude a genuine work partner.