Ways to get to a sharp answer
Every framework idea finishes the same way. You take your draft and run it through a Sharpen prompt until two people would make the same decision from it. Sharpen needs a draft to work on though. If you cannot write one yet, or it feels vague, these are ways to get to a draft worth sharpening. Whichever you use, you finish the same way, by running it through Sharpen.
When to use these
- You are staring at a blank page and cannot write the first draft.
- You have a draft but it feels vague and you cannot see why.
- You have too many ideas and cannot choose between them.
Interview me
Let AI ask you the questions. Instead of writing the draft yourself, you answer a short interview and AI assembles the first draft from your answers. Good for a blank page.
Copy this into your AI chat.
Replace the bracketed part with what you are defining, then answer its questions one at a time. It writes the first draft from your answers, and you run that through Sharpen.
Doing this with Claude Code: Interview me
Ask Claude to question you before it builds anything. One question at a time, broad first and then specific, until the gaps are closed. Skip this and it charges ahead on its own guesses and wastes time on the wrong thing. The payoff is bigger than Claude filling in what it needs, because the questions bring up things you had not thought of yourself, so your own thinking comes out more complete. That matters most when the area is new to you, where you do not know what you do not know. The questions raise the things you could never have asked, the ways it can fail, the edge cases, the thing that breaks under load.
The interview works in a few ways: one question at a time, broad then specific and the questions you would not have asked.
One question at a time. It asks, you answer, then the next. A barrage at once and you cannot think, one at a time and each answer shapes the question that follows.
Broad first, then specific. Start wide, who is this dish for, then narrow, any allergies, how much time have you got.
The questions you would not have asked. The real payoff is the things it raises that you did not know to think about, so your own thinking comes out more complete.
No drafting until it has asked. Tell it not to write the card until the questions are done, or it charges ahead on its own guesses.
Show by example
Give AI an example or two and let it infer the shape for yours. Good when you can recognise a good answer but cannot generate one from scratch.
Copy this into your AI chat.
Replace the bracketed parts and paste one or two definitions you think are good. It drafts one in the same shape for your case, which you then run through Sharpen.
Doing this with Claude Code: Show by example
Do not lean on a flowery description of what you want, show the result you expect instead. Plain words get read a little differently each time so the output drifts, and a concrete example removes that room and pins it down. Use it as a fallback first. Start by describing what you want, and if that keeps giving mixed results, stop rewording it and switch to an example. Begin with the happy path, the normal case working as it should, then build from there. Keep each one concrete and narrow, because a broad example pins nothing down.
Vary your examples rather than repeating one shape, because a few genuinely different ones teach the underlying pattern and Claude can then handle a case you never showed it, where near-identical ones only teach that single case. Show the reasoning as well as the result when the call is a judgment one, why you chose this over the obvious alternative, so Claude picks up the principle and not just one answer. If a term in your example is loose, define it by example so Claude works from your meaning and does not invent its own. When you want Claude to judge what is acceptable, show one acceptable case next to one real problem so it learns where the line sits and stops flagging the harmless variation.
It earns its keep on the edge cases too. When the code trips on the odd one, the null, the empty, the boundary, hand over that single case as a worked example, this exact input gives this exact output, and let Claude fix just that. It works the other way as well, when Claude does something you do not understand, ask it to show you with an example.
An example earns its keep in five ways: consistency, generalisation, exact shape, ambiguity and reasoning.
Consistency, hold the shape. Show one recipe card laid out as a title, then ingredients, then steps, and every card comes back in that same shape, run after run.
Generalisation, a dish you never gave it. Show a few genuinely different cards, a dinner, a dessert and a drink, and Claude can write a snack card it never saw, because the spread taught it the pattern and not one layout.
Exact shape, leave nothing to guess. Show one card with every field filled in, so the shape is there to copy rather than described.
Ambiguity, pin down the vague bits. Show what salt to taste means once, a pinch building to about half a teaspoon for this soup, and Claude reads the next loose instruction the same way.
Reasoning, the why and not just the result. Show why a pinch and not a spoonful, because more would overpower it, and Claude picks up the principle instead of copying one amount.
Give me options
Ask for several distinct candidates and choose. Good when you have a rough idea but cannot commit, or want to see the range before you pick.
Copy this into your AI chat.
Replace the bracketed parts with your own. It gives you four genuinely different candidates, then you pick the closest one and run it through Sharpen.
Doing this with Claude Code: Give me options
Ask Claude for several genuinely different options at once, not slight rewordings of one idea, and have it lay them side by side. Seeing the range is easier than generating the one right answer from a blank page. Pick the closest and tell Claude what to change, and you have a draft to sharpen.
Options work best when they are genuinely different and laid side by side, so you can see the range and then refine.
Genuinely different, not rewordings. Four real alternatives, a pasta, a stir fry, an omelette and a soup, not four versions of the same dish.
Side by side. Lay them together so you can compare, seeing the range is easier than naming the one right dish from nothing.
Pick the closest, then refine. You do not need a perfect option, just the nearest one, then tell it what to change to make it yours.
Then sharpen
Whichever on-ramp you used, you now have a draft. Finish the way every idea does: paste it into the Sharpen prompt on the idea you are working on, and tighten it until two people would make the same decision from it.
Say it another way
Sharpen is this site's name for the final step: you paste your draft into a set prompt that pushes you to make it clear and specific, until two people reading it would make the same decision. It is the difference between a rough note to self and a brief someone else could actually act on.
Doing this with Claude Code: Checkable criteria
A gap is only specific if you could test it, and if you cannot say how you would check the result the ask is still too vague. Vague in, vague out, the old garbage in garbage out (GIGO) where the garbage is vagueness. Claude still tries, but it wanders into work that feels like progress and fixes nothing. So make the criterion checkable and keep it to one small thing. Do not ask for a soft quality like be conservative or show the most important things, because important is your call and Claude cannot know it on its own. Name the dimension instead, performance or security, and if that is still broad go one level finer, so security becomes PII in logs. Sort by what is at stake, not by how sure Claude feels, by saying what to always report and what to skip. Pin any vague dial to a concrete example too, so critical is anchored to a real snippet and not a feeling. One noisy category and you start to become unsure of what to trust in the project, so when one goes noisy, mute it, fix it properly, then bring it back.
A criterion is sharp when it is checkable not vague, names the dimension, sorts by stakes and anchors severity. Mute the noisy category when it cries wolf.
Checkable, not vague. If you cannot say how you would test it, the ask is still too vague. Flag the bake only if the oven is outside 180 to 200 beats check the temperature is sensible.
Name the dimension, then go finer. Do not say be careful or the important bits, name what matters here, the seasoning, the timing, the food safety, and if that is still broad narrow it, food safety becomes anything served undercooked.
Sort by stakes, not confidence. Say what to always flag, anything unsafe, and what to skip, a wonky garnish, rather than asking it to guess what is important.
Anchor severity to an example. Critical, major and minor mean nothing on their own, pin each to a concrete case so it grades the same way every time.
Mute the noisy one. If one category cries wolf, turn it off, fix it properly, then bring it back, before its false alarms make you distrust the rest.
Doing this with Claude Code: Grouping your fixes
When you have a few fixes to give, decide how to group them first. Ask one question, do these affect each other. If fixing one would change the others, put them all in one detailed message with the whole picture of how they join up, so Claude can balance them together. A vague list hands the steering back to it. If the fixes are independent, send them one at a time and check each before the next, otherwise its fix for the next thing breaks the last one and it goes round in circles. Grouping the work this way is your job, not Claude's. And do not just hit return on what comes back. Stop, read what it did and if it has drifted somewhere you did not expect, reframe and ask questions until it is clear. It might be right, but you are the one driving.
How you send fixes depends on whether they affect each other, one message or one at a time. Then read what it did.
Ask if they affect each other. That one question decides how to send your fixes.
Dependent, one message. If changing the oven temperature also changes the timing, give both together with the whole picture, so it can balance them.
Independent, one at a time. A typo and the salt do not touch each other, so send them separately and check each, or its next fix undoes the last.
Read what it did. Do not just hit return, look at the change, and if it has drifted, reframe and ask questions. You are the one driving.
In plain words
In plain words: this page is about getting from a vague idea to a clear first draft, for when you are stuck at a blank page. It offers three ways to get AI to help: let it interview you, show it an example to copy, or ask it for a few options to choose from. Whichever you use, you then run the draft through Sharpen to tighten it. AI here is a tool you talk to in plain language, and Sharpen is this site's name for a prompt that pins a draft down until it is clear.
Part of Working with AI.
These on-ramps run across the Shape ideas.
The soft skills behind them: The soft skills of working with AI.
Why Claude. We use Claude Code in the notes above because it is one of the most widely used AI tools for building software. It is also what we build with day to day. The on-ramps themselves are general and carry across to other capable AI tools.