The problem this solves
Most disappointing AI output isn’t a model failure. It’s an under-specified request.
You type “make this page nicer” and get something generic. You type “write a script to clean up my data” and get code that guesses at your file format. The model wasn’t wrong, it answered the question you asked. The question just didn’t contain enough to answer well.
Prompt is the skill that fixes the question before you spend a run on the answer.
Worth being precise about scope: this is about the instructions you send to an AI. It has nothing to do with making your software faster or your code shorter.
What goes wrong without it
- Vague adjectives do the heavy lifting. “Clean”, “modern”, “professional”, “better”, these feel like requirements but carry almost no information. Two people reading them picture different things, and so does the model.
- The output arrives in the wrong shape. You wanted three bullet points and got four paragraphs, because nobody said what the answer should look like.
- Backstory crowds out the signal. Long preambles that don’t change the answer dilute the parts that do. More words is not more clarity.
- You find out it was wrong after the expensive part. A thin prompt on a large task means you discover the misunderstanding at the end, having paid for the whole detour.
How it actually works
It gets your real intent first. If the goal is genuinely ambiguous, MasterMind asks you one or two sharp questions rather than inventing requirements you never mentioned. This matters: the job is to sharpen your intent, not to replace it with a more convenient one.
Then it rewrites, applying only what fits:
- The task leads. Goal in the first line, plainly, with a clear picture of what “done” means.
- Only load-bearing context stays. The why, the audience, the constraints, what already exists. Everything that doesn’t change the output is cut.
- Vague becomes concrete. “Fast” becomes a number. “Nice” becomes a specific reference to match.
- The parts get separated. Context, task, requirements, and output format are labelled, so the model can tell instructions from background instead of guessing.
- An example appears when words aren’t enough. One short sample of the desired output usually beats three sentences describing it.
- The output contract gets pinned. Format, length, what to include, what to leave out.
Shorter is a side effect, not the goal. Padding and repetition go; every detail that changes the answer stays.
When it fires
You don’t need a command. Any of these reaches for prompt:
“can you improve this prompt before I run it?” “how should I ask for this?” “I keep getting bad answers from this: what am I doing wrong?” “here’s what I want to ask another AI, make it sharper”
You’ll see it engage in your terminal:
🧠 MasterMind ▸ sharpening this before you send it
└ prompt · intent → rewrite → what changed
When it does not fire
- You pasted a prompt for MasterMind to answer or run. This is the important one. A prompt in your
message is usually a task, not a rewrite request, so
promptstays out of the way unless you actually ask for it to be improved. Rewriting text you meant to be executed would replace your work with a paraphrase and answer a question you never asked. - You want your code optimized. That’s
performance, measuring and fixing slow software.promptnever touches your code; it only touches text you send to a model. The names sound adjacent; the jobs share nothing. - The ask is fuzzy but it’s a build task, not a prompt. If you want MasterMind itself to build
something and the requirements are unclear, that’s
interview, it turns a vague ask into a buildable specification.promptis for when the finished text is the deliverable and you’ll send it somewhere else. - You want to teach MasterMind a lasting preference. That’s
levelup. A rewritten prompt is a one-off artifact; it doesn’t change future behavior.
What you get
Three things: the rewritten prompt in a block you can copy straight out, two to four bullets explaining what changed and why, and an honest list of anything still ambiguous that only you can decide.
Scope is never silently expanded. If sharpening the request surfaces a real product decision, MasterMind puts it in front of you rather than quietly picking for you.
Your prompt stays yours. The rewrite is handed back, never run. MasterMind doesn’t execute it, doesn’t overwrite your original, and doesn’t act on it in the same turn. Every requirement you wrote survives or is named as cut, and anything added is named as added, so you can always see exactly what changed. That rule isn’t local to this skill: it binds every skill that touches your own words: your spec, your style profile, your docs.