Prompt Engineer

Prompt Engineer

Purpose

Design production-grade prompts: system instructions, task templates, variable slots, few-shot examples, and output format contracts.

Input: Task definition, AI architecture (optional), example inputs/outputs (optional), constraints Output: Prompt specification with template, variables, examples, and test cases References: See [references.md](references.md) for domain cheat sheets. Examples: See [examples.md](examples.md) for worked input/output.

---

Workflow

Step 1: Define prompt objective

Specify:

• Single task the prompt performs (one objective per prompt)
• Success criteria (what good output looks like)
• Failure modes to prevent

Step 2: Structure the prompt

Use layered structure:

1. Objective — what to accomplish
2. Context — background information slots
3. Instructions — step-by-step task rules
4. Constraints — what not to do
5. Output format — exact structure expected
6. Examples — few-shot demonstrations (2–3)

Step 3: Define variables

| Variable | Type | Required | Description | |----------|------|----------|-------------|

Use {{variable_name}} placeholders. Document validation rules.

Step 4: Write output format contract

Specify exact output structure:

• JSON schema description, or
• Markdown sections, or
• Structured text pattern

Include handling for edge cases (empty input, ambiguous input).

Step 5: Create test cases

| ID | Input | Expected behavior | Pass criteria | |----|-------|-------------------|---------------|

Minimum 5 test cases: happy path, edge cases, adversarial input.

Step 6: Validate

Run Validation checklist.

---

Decision Rules

| Condition | Action | |-----------|--------| | Multiple unrelated tasks in one prompt | Split into separate prompts | | No output format specified | Define strict format before delivery | | Task requires external data not in variables | Add variable or recommend RAG/tool integration | | Prompt exceeds ~2000 tokens template | Modularize; use context-engineering skill | | Customer-facing without refusal behavior | Add constraint block for out-of-scope requests |

---

Validation

• [ ] Single clear objective stated
• [ ] All 6 prompt layers present
• [ ] Variables documented with types and validation
• [ ] Output format contract is unambiguous
• [ ] ≥2 few-shot examples included
• [ ] ≥5 test cases with pass criteria
• [ ] No Role Play or persona assignment phrasing
• [ ] Refusal/out-of-scope handling for user-facing prompts

---

Anti-patterns

• Mega-prompt — everything in one undifferentiated block.
• Vague instructions — "be helpful", "be accurate" without specifics.
• No output format — free-form when structure is needed.
• Missing negative constraints — not saying what to avoid.
• Example pollution — examples that contradict instructions.

---

Best Practices

• Put instructions after context (recency bias).
• Use delimiters for variable content (---, XML tags).
• Test with edge cases before production.
• Version prompts with changelog.
• Pair with prompt-reviewer before deployment.

---

Output Structure

# Prompt Specification: [Prompt Name]
**Version:** 1.0

## Objective
[Single task description]

## Template

[Full prompt with {{variables}}]

## Variables
| Name | Type | Required | Validation |
|------|------|----------|------------|

## Output Format
[Exact structure]

## Examples
### Example 1
**Input:** ...
**Output:** ...

## Test Cases
| ID | Input | Pass criteria |
|----|-------|---------------|

## Changelog
| Version | Change |
|---------|--------|

---

Next Skills

| Outcome | Recommended Skill | |---------|-------------------| | Review prompt quality | ai/prompt-reviewer | | Optimize token usage | ai/prompt-optimizer | | Manage context budget | ai/context-engineering | | AI architecture context | ai/ai-solution-architect |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: patonkikh
• Source: patonkikh/APES
• License: MIT

Install and usage instructions live in the source repository linked above.