| name | prompt-engineering-expert |
| description | Expert prompt engineering, custom instruction, system prompt, and agent instruction design. Use when Codex needs to review, generate, refactor, debug, optimize, document, or test AI prompts; design reusable prompt templates; create or improve system prompts, custom instructions, agent behavior guidelines, tool-use prompts, multimodal prompts, or prompt evaluation frameworks.
|
| license | MIT |
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.
Custom Instructions And Agent Prompts
When designing system prompts, custom instructions, or agent skills:
- Define the agent's role through responsibilities and decision standards, not theatrical persona.
- Separate mandatory behavior from style preferences.
- Include boundaries: what to refuse, what to escalate, what to ask about, and what to infer.
- Keep instructions stable across turns; avoid directions that require mutating past context.
- For tool-using agents, specify tool selection, validation, retry, and user-update behavior.
- For coding agents, include repository conventions, test expectations, and change-safety rules.
Anti-Patterns
Watch for:
Vague verbs: "analyze," "improve," "make better," "handle this."
Contradictions: "be concise" plus many mandatory sections, or "do not ask questions" plus missing required data.
Overfitted examples that teach accidental details.
Output formats described in prose when a schema or example is needed.
Prompts that invite hallucination by asking for facts without sources or data.
Security gaps: untrusted user content can override instructions, leak context, or request unsafe actions.
Token bloat from background essays, duplicated rules, and unused options.
See references/troubleshooting.md [blocked] for failure modes and fixes.
Evaluation
Every non-trivial prompt improvement should include tests:
- Happy path with typical input.
- Edge case with missing, ambiguous, or malformed input.
- Regression case for a known failure.
- Adversarial or injection case when untrusted input is involved.
- Format compliance case when downstream parsing matters.
- See references/evaluation.md [blocked] for test templates and scoring rubrics.
Reference Loading
Load only the references needed for the request:
- references/techniques.md [blocked]: prompting techniques, when to use each, and compact examples.
- references/troubleshooting.md [blocked]: common prompt failures, diagnosis, and fixes.
- references/evaluation.md [blocked]: prompt test cases, rubrics, and regression strategy.
- references/examples.md [blocked]: reusable prompt review, generation, classification, structured-output, and agent-instruction examples.
