Prompt patterns that actually reduce rework

Why vague prompts create thrash

If you write “build me a CRM,” the model invents a product: random dashboards, fake metrics, auth you did not ask for, and libraries you cannot explain. You then spend hours deleting.

A better prompt is slightly longer up front and much shorter overall — because you do not re-generate the whole app five times.

The three-layer pattern

Layer 1 — Context: who you are, skill level, audience, and what “success” means this session.

Layer 2 — Scope: concrete screens, data, and behaviours. Prefer lists of must-haves over vibes alone.

Layer 3 — Constraints: must-nots. Security, no fake stats, accessibility, stack limits, “explain new dependencies.”

You will see this pattern across the prompt library on this site, including game one-shots (where “polish bar” is a kind of constraint).

Bad vs better (short example)

Bad: “Make a habit app with AI and make it beautiful.”

Better: “Context: I am a beginner; personal tool only. Scope: add habits, mark complete today, show a streak number, save in the browser. Constraints: no accounts, no paid APIs, accessible labels, mobile layout, do not invent social features.”

The better version still leaves room for the model to design UI — it just closes doors you would have slammed later.

Iteration prompts that help

After the first draft, avoid “make it better.” Try:

• “Do not add features. Fix only: empty state, mobile overflow, and form labels.”
• “Explain in five bullets what you changed and why.”
• “Point out anything insecure you introduced; patch the top two only.”
• “Reduce this to one file if possible; remove unused dependencies.”

One-shots vs multi-step

One-shot prompts (like Sky Dash) try to ship a complete toy in one go. They work best for closed worlds: games, calculators, single-page demos.

Multi-step is safer for anything with logins or money: vertical slice first (one path works), then harden, then polish.

If a one-shot fails, do not re-roll forever. Keep the best half, and prompt for the missing milestone only.

Copy-paste starter skeleton

CONTEXT: … (who, skill, audience)

SCOPE: … (screens + behaviours)

CONSTRAINTS: … (must-nots + quality bar)

DONE WHEN: … (how you will judge success in 30 minutes)

Steal filled versions from the Prompts library; adapt the context line to you.

Related

Get started walks a first session. How to test prompts shows how to judge output. When to stop accepting diffs covers the moment review must win over speed.