Prompt Improver
Paste a weak prompt, get a structured one back.
- $25 Free
- 30 sec
- No signup
Paste your current prompt
See it scored and rewritten
Copy the improved prompt into the CLI
You get: A rewritten, structured prompt plus what was fixed and why.
Your prompt
What it checks
- ✗Specific, not one vague verb
- ✗Includes context / the why
- ✗Points at files or components
- ✗Defines what 'done' looks like
- ✗Bounds the scope
- ✗Asks for a specific output
- ✗Asks for a plan first (for big tasks)
Improved prompt
## Task fix the login bug ## Context [Why this matters + what the code currently does + anything already tried] ## Where [Name the file(s), component(s), or path(s) to work in] ## Done when [The concrete condition that means success - a passing test, correct output, no errors] ## Scope / constraints [What to change - and explicitly what NOT to touch] ## Output [The shape you want back - e.g. reply with only the diff, or a short bullet list] ## First Propose a short plan before writing any code. Wait for my go-ahead if the approach is non-obvious.
What was missing
Specific, not one vague verb
A prompt like 'fix the bug' gives Claude nothing to work with. Say what is broken, where, and what correct looks like.
Includes context / the why
Add the surrounding situation: what the code does, why you want the change, what you have already tried.
Points at files or components
Name the file, path, function, or component. Claude Code works far better when it knows where to look.
Defines what 'done' looks like
State the acceptance criteria - 'the test passes', 'the endpoint returns 200', 'no console errors'. Without a target, you get a guess.
Bounds the scope
Say what NOT to touch. 'Only change the auth handler, leave the routes alone' prevents unwanted sprawl.
Asks for a specific output
Tell it the shape you want back - a diff, a JSON block, a short list. This saves tokens and matches your workflow.
Asks for a plan first (for big tasks)
For anything non-trivial, ask Claude to propose a plan before writing code. Catching a wrong direction early is cheap.
The rewrite is a scaffold - fill the bracketed sections with your real details before sending. The structure is what makes Claude answer well; you still supply the facts.
Why most prompts underperform
The gap between a mediocre Claude Code result and an excellent one is almost never the model - it is the prompt. A vague prompt like 'fix the login bug' forces Claude to guess at the context, the file, and what 'fixed' even means. Those guesses are where wrong directions and wasted tokens come from. This tool reads your prompt against the handful of traits that reliably separate strong prompts from weak ones, scores it, and hands you a structured rewrite with the gaps called out.
The analysis is a set of transparent heuristics running entirely in your browser - it checks for the presence of context, file references, success criteria, scope bounds, output format, and a plan request. It is not judging your writing style; it is checking whether the ingredients a good coding prompt needs are actually in the text.
The six ingredients of a strong coding prompt
- Specificity - name the actual problem, not a one-word verb. 'Fix' tells Claude nothing; 'the login form accepts an empty password' tells it everything.
- Context - the why and the current state. What does the code do now, why do you want the change, what have you already tried.
- Location - the file, path, component, or function to work in. Claude Code is dramatically better when it knows where to look.
- Success criteria - the concrete condition that means done. A passing test, a 200 response, no console errors.
- Scope - what to change and, crucially, what NOT to touch. This is the single best defense against unwanted sprawl.
- Output format - the shape you want back. 'Reply with only the diff' saves tokens and matches how you actually work.
Scope is the most underused lever
Adding one line - 'only change the auth handler, leave the routes and tests alone' - prevents most of the over-eager edits that make people distrust AI coding. Telling Claude what to leave alone is as valuable as telling it what to do.
Ask for a plan before code on anything non-trivial
For a one-line fix, just ask for the fix. For anything with more than one moving part, ask Claude to propose a short plan first and wait for your go-ahead. A 200-token plan that catches a wrong approach is far cheaper than a 5,000-token implementation headed in the wrong direction. This is both a quality lever and a cost lever - it is the highest-leverage habit in Claude Code and it costs almost nothing.
Structure beats prose
The rewrite this tool produces uses headed sections - Task, Context, Where, Done when, Scope, Output - because a structured prompt is easier for the model to parse and harder for you to leave incomplete. You do not need to be a good writer to prompt well; you need to include the right pieces. The scaffold guarantees you do. Fill in the bracketed placeholders with your real details and you have a prompt that consistently outperforms a paragraph of stream-of-consciousness.
What this tool does not do
It does not call any model. There is no API key, no network request, no data leaving your browser. That means it cannot judge whether your facts are correct or your approach is sound - only whether the prompt contains the structural ingredients that make Claude answer well. Think of it as a checklist enforcer, not an oracle. The judgment about the actual task is still yours; the tool just makes sure you did not skip a section that would have cost you a bad answer.
Turning this into a habit
- Before sending any non-trivial prompt, glance at whether it has context, a location, and a success criterion. Those three fix the majority of bad results.
- Keep a snippet of the scaffold headings handy and paste it as a template for real tasks.
- When a result disappoints, re-read your prompt first - nine times out of ten the gap was in the ask, not the model.
- Pair a scoped prompt with a plan request for large changes; ship small, verify, then continue.
Frequently asked questions
Does this send my prompt to Claude or any server?
No. The analysis and rewrite happen entirely in your browser using transparent heuristics. Nothing you type is uploaded, logged, or sent anywhere - the copy button just puts the improved prompt on your clipboard.
How is the score calculated?
It checks your text for six ingredients of a strong coding prompt - specificity, context, file references, success criteria, scope bounds, and output format - plus a plan request. The score is the share of those that are present.
The rewrite has bracketed placeholders. Do I send it as-is?
No - fill the bracketed sections with your real details first. The tool provides the structure; you supply the facts. Sending it with placeholders still in would give Claude an incomplete brief.
Why does it flag my prompt when it looks fine to me?
It is checking for structural ingredients, not readability. A prompt can read clearly and still omit the file to edit or the definition of done - which is exactly what leads to a good-looking answer that misses the point.
Is a longer prompt always better?
No. The goal is completeness, not length. A tight prompt that names the file, the change, and the success condition beats a rambling paragraph. Add detail where it removes ambiguity, not for its own sake.
Does this replace learning to prompt well?
It accelerates it. By showing you what was missing every time, it trains the instinct to include those pieces yourself. After a while you will write structured prompts without the tool - which is the point.
Liked this tool? The club is the next step.
Join Claude Code Club for $9/month. 650+ lessons, weekly updates, and the workflows behind every tool on this site.
- No experience needed
- Cancel anytime
- Updated weekly