prompt-craft-en

Prompt Craft — Prompt Design & Optimization

Turn the user's intent into a clear, reusable, ready-to-run prompt. The methodology is a field-tested four-element framework: Goal, Context, Output, Boundaries.

A good prompt needs no special syntax or fixed template. Add elements only when they help — not every one, every time.

---

Workflow

When the user asks you to write or optimize a prompt, follow these steps.

1. Identify the task type

Type	Traits	Template
Everyday	A question, brainstorming, a draft, a comparison, a small plan	Short prompt, 1–3 sentences
Deliverable	Multiple sources/steps, produces a file, shown to others	Full four elements + final check
Code	Working on code, fixing a bug, writing tests, refactoring	Behavior + repro/location + constraints + how to verify
Agent / system	System prompts, automation, scheduled runs	Four elements + approval boundaries + failure handling

2. Fill gaps quickly

If the user hasn't given enough to pin down a key element, ask at most one most-important question (Audience? Purpose? What must not change?). When there's basically enough, just start — write your assumptions as annotations next to the prompt instead of asking round after round.

3. Produce the prompt

• Put the final prompt in a code block for easy copying.
• Write the prompt in the user's target language (English for English models/contexts; otherwise the user's language).
• Add 2–4 sentences of notes: why it's organized this way, where it can be trimmed or adjusted.
• For complex tasks, offer two versions: a lean one and a full one.

4. When optimizing an existing prompt

Diagnose before rewriting. Check for the common ailments, in priority order:

1. Unclear result: lots of steps, but no statement of what to produce or for whom → rewrite around the result.
2. Context noise: irrelevant sources/background stuffed in → keep only what changes the result.
3. Missing boundaries: doesn't say what's off-limits or what needs approval → add the 1–2 that matter most.
4. Over-control: every step nailed down, no room to adjust → delete procedural instructions, keep the constraints.
5. No verification: an important task with no final check → add a closing self-check.