Say It Like a Human

The request structure that gets finished things, not fragments.

By the end of this lesson

You'll write requests that let Claude work on its own for long stretches.

You don't need magic words. You need a complete request — everything a smart helper would need to finish the job without coming back to ask. Five simple ingredients.

First: four chat habits to unlearn

1Drip-feeding tiny questions. In a chat you ask piece by piece and assemble the answers yourself. Here, you hand over the WHOLE job — Claude does the assembling.
2New chat for every question. In a project, stay in ONE session: it holds the conversation's full memory. (A fresh chat loses the conversation, not your files — Claude can re-read the folder; you just waste time re-orienting it.)
3Hitting 'regenerate' when you don't like the answer. Don't reroll — steer. Say exactly what to change ('darker', 'bigger buttons') and it changes that, keeping the rest.
4Copy-pasting results out. There's nothing to copy anymore. The result isn't text in a window — it's real files appearing in your folder, already working.

The five ingredients

1Goal — the real result you want, said as an outcome.
2Context — who it's for, any constraints, which files.
3Examples — one or two of 'good', so it can match a target.
4Standards — what 'done' means: style, quality bar, what to avoid.
5Autonomy — explicit permission: 'make reasonable decisions and keep going; don't stop to ask.'

Weak vs. strong

weak   ▸ make me a website
strong ▸ Build a one-page site for a wedding photographer named Mara.
         Calm, lots of white space, one big hero photo, an about
         paragraph, a gallery of 6, a contact section. Make smart
         choices, build the whole thing, then tell me how to preview it.
         Don't stop to ask me questions.

→ the strong prompt returns a finished site. the weak one returns a question.

ℹShow, don't only tell — your screenshot habit works here

Used to attaching screenshots in ChatGPT? Same thing works in the terminal: drag an image file straight into the Claude Code window (or paste its file path) — Claude sees images. One catch: it needs a FILE. On a Mac, screenshots land on the Desktop as files, ready to drag; on Windows, Win+Shift+S only copies to the clipboard — save it first (the Snipping Tool preview has a save button). 'Make it look like this' + a screenshot is one of the strongest requests there is.

✦The autonomy line is magic

Telling Claude 'use your best judgment and build the whole thing before checking back' unlocks long, productive runs. Most people accidentally leash it with vague, timid prompts.

Author

Evgeny Arsentyev

PhD · Chief Product Officer at a healthtech company

About the author →