Turns a vague request into a structured, reliable prompt - role, context, task, format, failure handling, and few-shot examples - that produces consistent output across real inputs. Use when someone asks "why does my prompt give inconsistent results", "write a prompt for this task", "the model keeps breaking my JSON", "how do I stop prompt injection from user input", or is building any LLM feature whose prompt was written ad hoc. Do NOT use for converting a working prompt into a reusable agent skill - use prompt-to-skill instead; do NOT use for measuring whether a prompt change improved quality - use llm-evaluation instead.
Click to play with sound.
---
name: Prompt Engineer
description: Turns a vague request into a structured, reliable prompt - role, context, task, format, failure handling, and few-shot examples - that produces consistent output across real inputs. Use when someone asks "why does my prompt give inconsistent results", "write a prompt for this task", "the model keeps breaking my JSON", "how do I stop prompt injection from user input", or is building any LLM feature whose prompt was written ad hoc. Do NOT use for converting a working prompt into a reusable agent skill - use prompt-to-skill instead; do NOT use for measuring whether a prompt change improved quality - use llm-evaluation instead.
---
# Prompt Engineer
Turn a vague request into a prompt that produces consistent, high-quality output on inputs you have not seen yet. The costly mistake this prevents is the adjective prompt: "be detailed and professional" reads like an instruction but constrains nothing, so output quality varies run to run and the author iterates blind. Structure and examples constrain; adjectives decorate.
## Operating procedure
Ordering inside the prompt is load-bearing, not stylistic: instructions come before context so the model reads the data already knowing what to do with it, and the most important constraint is restated last because recency makes it stick.
### Step 1: Gather inputs
1. The single objective, stated in one sentence. If it takes two sentences, it is two prompts - split into a chain.
2. 3-5 real example inputs, including one ugly one. Iterate on real inputs, not imagined ones; label synthetic examples as synthetic.
3. Who or what consumes the output: a downstream parser wants JSON with a schema; a human wants markdown.
4. The failure policy: what the model should do when input is missing, ambiguous, or out of scope (default: return "UNKNOWN" rather than guess).
### Step 2: Build the four blocks, in this order
```text
ROLE: You are a senior <domain> expert.
TASK: <imperative, one goal - placed before the context it operates on>… load the full skill through Skill Me