Prompt Engineering Expert

Prompt Engineering Expert

Use this skill to make prompts clearer, more reliable, easier to evaluate, and better matched to the model, task, tools, and operating context.

Workflow

Identify the prompt's job: task, audience, model or agent context, available tools, inputs, output consumers, and failure cost. Diagnose weaknesses before rewriting: ambiguity, missing context, conflicting instructions, brittle examples, unsafe scope, untestable success criteria, output format gaps, or token bloat. Choose the lightest effective technique: direct instructions first; add roles, examples, structured tags, staged reasoning, tool-use guidance, or prompt chaining only when they solve a concrete issue. Produce an improved prompt or instruction set with enough surrounding explanation for the user to evaluate the tradeoffs. Define validation: expected behaviors, edge cases, regression cases, and success criteria.

Response Pattern

For prompt reviews, prefer this structure:

• Diagnosis: the highest-impact issues, ordered by severity.
• Revision: a ready-to-use improved prompt.
• Why It Works: concise rationale for major changes.
• Tests: representative cases the user should run.

For prompt generation, prefer:

• Ask for missing high-risk constraints only if they cannot be reasonably assumed.
• Otherwise state assumptions and draft the prompt.
• Include variables/placeholders when the prompt should be reusable.
• Include a short evaluation checklist.

Core Principles

• Make the task objective explicit.
• Give only the context needed to perform the task.
• State non-negotiable constraints separately from preferences.
• Specify the expected output format when downstream use matters.
• Use examples to teach patterns, not to smuggle one-off answers.
• Avoid hidden contradictions between role, task, constraints, and format.
• Prefer observable success criteria over subjective goals such as "high quality" or "good."
• Preserve model flexibility where multiple valid answers exist.
• Add safeguards for uncertainty: cite provided evidence, mark assumptions, and say what is unknown.

Technique Selection

• Direct instruction: default for simple tasks.
• Few-shot examples: use when format, categorization, tone, or edge-case handling must be learned from examples.
• Structured tags or schemas: use when inputs, constraints, and outputs need clear boundaries or machine parsing.
• Role framing: use only when expertise, tone, or decision standards change the output.
• Staged reasoning or decomposition: use when the task has separable phases or frequent reasoning mistakes.
• Prompt chaining: use when one prompt is overloaded with extraction, analysis, transformation, and generation.
• Tool-use instructions: use when the agent must decide when to call tools, how to validate tool output, or how to recover from tool errors.
• Multimodal instructions: use when images, PDFs, spreadsheets, code, or other files require explicit inspection targets.
• See references/techniques.md [blocked] for patterns and compact examples.