What is a prompt optimizer?
A prompt optimizer is a free tool that turns a rough, half-formed prompt into a clear, structured one that gets far better results from any AI. Paste your rough idea, and this free prompt optimizer rewrites it with the parts that actually matter - a role for the AI, the task, the context, your requirements, and the output format you want. The difference between a vague request and a structured prompt is the difference between generic and genuinely useful output.
It also works as a prompt enhancer and ChatGPT prompt optimizer: the structured prompt it produces is portable, so you can paste it straight into Claude, ChatGPT, Gemini, or any AI chat. No prompt-engineering knowledge needed - you describe what you want in plain language and the optimizer handles the structure.
Need a prompt from scratch? Switch to Generate a prompt above: describe the goal in one sentence and the prompt generator writes a complete, reusable prompt with a role, step-by-step instructions, constraints, an output format and [placeholders] for the details only you know.
Want to skip prompting entirely for recurring work? A ready-made prompt pack or skill file from KissMySkills turns Claude or ChatGPT into a specialist for a specific job, so you get expert output without writing prompts at all. New to prompting? Read what Claude Skills are to see how a saved skill beats re-prompting every time.
Where you put things in the prompt changes the answer
Position is the least discussed and most measurable part of prompt structure, and two of the three big model vendors publish the same advice.
Anthropic's prompting guidance, for inputs in the 20,000-token range and above, tells you to "place your long documents and inputs near the top of your prompt, above your query, instructions, and examples," and states that queries at the end "can improve response quality by up to 30 percent in tests, especially with complex, multidocument inputs." Treat that percentage as an in-house figure without published methodology, but the ordering advice itself is unambiguous. Google gives the same instruction for Gemini: "When providing large amounts of context (e.g., documents, code), supply all the context first. Place your specific instructions or questions at the very end of the prompt," and adds the useful bridging trick of a transition phrase such as "Based on the information above..."
There is research behind the intuition. Lost in the Middle (Liu et al., arXiv:2307.03172) found that model performance "is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models." If the fact you need is buried halfway through a pasted document, moving it is a more reliable fix than rewording the question.
OpenAI's current guidance orders a developer message differently - Identity, Instructions, Examples, then Context, with the note that context "is usually best positioned near the end of your prompt." That is not a contradiction so much as a difference in what is being placed: request-specific reference material versus bulk documents. The practical takeaway is that there is no single universal layout, and a prompt tuned by position on one model should be re-checked on another.
Structure beats adjectives
The techniques the vendors actually document are structural, and none of them involve telling the model it is a world-class expert.
Anthropic's framing for clarity is a test rather than a rule: "Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too." The same guidance recommends explaining why a constraint exists rather than stating it flatly, on the grounds that the model "is smart enough to generalize from the explanation" - an instruction with a reason attached transfers to cases you did not enumerate.
Delimiters are the other consistent recommendation. Anthropic: XML tags "help Claude parse complex prompts unambiguously," and wrapping each type of content in its own tag such as <instructions>, <context>, <input> "reduces misinterpretation." OpenAI: "Markdown headers and lists can be helpful to mark distinct sections of a prompt," and "XML tags can help delineate where one piece of content... begins and ends." Google warns in the other direction about examples: consistent formatting across few-shot examples matters, and "if you include too many examples, the model may start to overfit the response to the examples." Anthropic suggests three to five, wrapped in <example> tags.
Roles do less than the internet implies but are not useless. Anthropic's position is modest and specific: "Setting a role in the system prompt focuses Claude's behavior and tone for your use case. Even a single sentence makes a difference." One sentence in the system prompt - not a paragraph of invented biography in the user message.
Two pieces of standard prompt advice that are now wrong
Prompt guides age badly, and two of the most-repeated tips have been overtaken by the vendors' own documentation.
"Always tell it to think step by step." OpenAI's reasoning guidance says the opposite for reasoning models: "Avoid chain-of-thought prompts: Since these models perform reasoning internally, prompting them to 'think step by step' or 'explain your reasoning' is unnecessary," and elsewhere that such techniques "may not enhance performance (and can sometimes hinder it)." The same page recommends trying zero-shot first and adding examples only if needed. OpenAI's analogy is worth keeping: a reasoning model "is like a senior co-worker. You can give them a goal to achieve and trust them to work out the details. A GPT model is like a junior coworker." Chain-of-thought is not dead - Anthropic still documents manual CoT with <thinking> and <answer> tags - but it is now described as a fallback for when built-in reasoning is off, not a default.
It is also worth knowing what the original paper claimed. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (Wei et al., arXiv:2201.11903) reported that such abilities "emerge naturally in sufficiently large language models." The famous "let's think step by step" result (Kojima et al., arXiv:2205.11916) was measured on a 2022 model. Neither is a timeless law.
"Prefill the assistant's first words." Anthropic's current documentation states that starting with recent Claude models, prefilled assistant messages on the last turn "are no longer supported," and requests containing them "return a 400 error." The documented replacements are structured outputs, a direct instruction to respond without preamble, or tool calling. Any guide still teaching prefill as a formatting trick is describing a request that now fails.
One more distinction worth having: self-consistency (Wang et al., arXiv:2203.11171) is a decoding strategy that samples several reasoning paths and takes the most common answer. It cannot be delivered by rewriting a prompt, because it requires multiple generations.
What this optimizer can't do
It can't supply knowledge the model doesn't have. OpenAI's stated reason for adding context is to give the model "access to proprietary data, or any other data outside the data set the model was trained on." The fix for a missing fact is pasting the fact, not phrasing the question better.
It can't eliminate hallucination. Anthropic says plainly that even advanced models "can sometimes generate text that is factually incorrect," and that its mitigation techniques - allowing "I don't know," extracting word-for-word quotes before answering long documents, restricting the model to supplied sources - "significantly reduce hallucinations" but "don't eliminate them entirely."
It can't protect you from prompt injection. Anthropic names two threat models: direct injection where the user is the adversary, and indirect injection where a trusted user's prompt causes the model to process third-party content - web pages, emails, documents, tool results - carrying adversarial instructions. Wording is not a security control.
It can't set your sampling parameters. Temperature and top-p live in the API call, not the prompt. Note they are not comparable across vendors either: Anthropic documents a 0.0-1.0 range and warns that "even with temperature of 0.0, the results will not be fully deterministic," while OpenAI documents 0-2 and recommends "altering this or top_p but not both."
It can't tell you whether the output is good. Anthropic's prompt-engineering docs assume you arrive with a definition of success and a way to test against it. Without that, a longer, better-structured prompt just produces a longer, better-structured guess.
