Prompting is editing, not incantation

The biggest unlock with AI is realising your first prompt is a draft, not a spell. You are not hunting for the one magic sentence that produces a perfect answer. You are starting a short conversation, reading what comes back, and telling the model exactly what to change. That feedback loop is the actual skill.

People who get great results from AI are rarely better at writing prompts. They are better at editing them. They look at a so-so answer and say “good start, now cut the intro and make it more casual”, and they get there in two or three rounds.

Why does the “magic words” idea fail?

Because no single sentence can capture everything you want before you have seen an attempt. Half of what you care about only becomes obvious once the model shows you a version that gets it slightly wrong. “Oh, too formal.” “Oh, I wanted bullet points.” That reaction is information, and the only way to get it is to see a draft first.

Chasing the perfect one-shot prompt also wastes time. You sit there polishing a paragraph of instructions when you could have sent a rough version, seen the gap, and fixed it in ten seconds of feedback.

What does the loop look like?

It is simple and it repeats.

Draft

Send a rough, honest version of what you want. Do not overthink it. You just need something on the page to react to.

Say what is wrong

This is the heart of it. Be specific. Not “make it better” but “the tone is too stiff, drop the second paragraph, and keep the examples shorter”. Concrete feedback gives the model something it can actually act on.

Refine

Let it revise based on your notes. Often one or two rounds is enough. If a change made things worse, just say so and roll it back. You are the editor, the model is the fast typist.

How do you give better feedback?

Treat it like editing a person’s draft. A few habits help:

• Point at the exact thing: “the opening line”, “the third bullet”, “the closing sentence”.
• Say the direction: shorter, warmer, more formal, more concrete, fewer adjectives.
• Keep what works: “keep the structure, just change the tone” saves the model from undoing good parts.
• Show, do not only tell: paste a sentence in the style you want and say “more like this”.

If you want the parts that make a strong starting draft, I broke those down in the anatomy of a good prompt. And if you are still adding “you are an expert” to everything, stop role-playing your prompts explains why that does little now.

The takeaway

Stop searching for the perfect prompt. Start a draft, read it like an editor, and tell the model what to fix. The people who seem to have a magic touch with Claude or any assistant are just running this loop quickly and without precious feelings about their first attempt. Prompting is editing. That is the whole trick.