prompt-engineer

prompt-engineer 3388 words MIT v1.2.0 wyattowalsh opus

Prompt engineering. Craft, analyze, harden, convert, design tool prompts, and build PromptOps/eval plans. Use for system, agent, tool, RAG prompts. NOT for running prompts or building agents.

Quick Start

Install:

npx skills add github:wyattowalsh/agents --skill prompt-engineer -y -g --agent claude-code --agent codex --agent gemini-cli --agent antigravity --agent github-copilot --agent opencode

Use: /prompt-engineer <mode> [target]

Works with Claude Code, Gemini CLI, Codex, and other agentskills.io-compatible agents.

What It Does

prompt-engineer designs and audits the full prompt context, not only the wording of a system prompt. It covers prompt structure, target-model fit, trust boundaries, tool definitions, eval design, and production PromptOps controls.

Current-source discipline

Model-specific guidance is time-sensitive. The skill records provider facts with evidence classes and last-verified dates, and flags stale guidance before recommending model-specific rollout changes.

Modes

$ARGUMENTS	Action
craft <description>	Craft a new prompt from scratch
analyze <prompt or path>	Diagnose and improve an existing prompt
audit <prompt or path>	Read-only prompt review
convert <source-model> <target-model> <prompt or path>	Port between model families
evaluate <prompt or path>	Build an eval framework
harden <prompt or path>	Stress-test injection, trust boundaries, and robustness
tool <tool definition or schema>	Design or review model-facing tool definitions
promptops <prompt or path>	Plan versioning, rollout, monitoring, and rollback
Raw prompt text	Auto-detect existing prompt and audit
Natural-language prompt request	Auto-detect craft mode
Empty	Show mode menu

Shared Preflight

All modes start with the same checks:

1. Ingest — Read prompt text or the supplied file path.

2. Classify — Identify mode, target model/provider, model class, deployment context, and trust boundary.

3. Load selectively — Read only the reference files needed for the selected mode.

4. Evidence-scope claims — Classify non-obvious recommendations as official docs, provider guide, research, community heuristic, local practice, or single-study.

5. Refuse wrong scope — Do not run prompts, build agents, or perform implementation work from this skill.

Production Modes

Harden

Use harden for prompts that handle user input, retrieved documents, tool results, external content, user-facing output, or write-capable tools. The output maps attack surfaces, controls, residual risk, and required adversarial evals.

Tool

Use tool for function schemas, MCP tools, OpenAPI fragments, or model-facing tool-selection guidance. The review checks name clarity, description specificity, parameters, errors, permission class, adjacent-tool overlap, and tool-result injection risk.

PromptOps

Use promptops to turn prompt changes into a deployable workflow: versioning, linked evals, rollout stages, monitoring, provider fact refresh, and rollback.

Model-Class Detection

Model Class	Examples	Guidance
Instruction-following	GPT-4.1-style, non-thinking Gemini/Claude modes, standard Llama variants	More explicit structure and examples can help; test before adding complexity
Reasoning	GPT-5.5/GPT-5.x reasoning modes, Claude extended thinking, Gemini thinking, o-series models, reasoning Llama variants	Avoid mandatory hidden CoT transcripts; use provider-supported reasoning controls or concise planning/preambles when current docs support them

Critical Rules

1. Do not recommend mandatory hidden chain-of-thought transcripts for reasoning models.
2. Use clear delimiters for multi-section prompts; XML is a strong default for Claude and many complex prompts, but provider guidance wins.
3. Run security review for any prompt handling untrusted input.
4. Recommend evals for any non-trivial prompt.
5. Cite reference files and last-verified dates for provider-specific claims.
6. Keep audit report-only and read-only.
7. Refuse and redirect requests to run prompts, build agents, or do non-prompt implementation work.
8. Update evals when dispatch behavior or modes change.
9. Run scripts/validate-references.py after reference, dispatch, or eval changes.

---

General

Field	Value
Name	prompt-engineer
License	MIT
Version	1.2.0
Author	wyattowalsh

Claude Code

Field	Value
Model	opus
Argument Hint	<mode> [target]

Related Skills

skill-creator Create, improve, and audit AI agent skills.

research Deep multi-source research with confidence scoring.

docs-steward Maintain generated and hand-maintained docs surfaces.

Source

View SKILL.md Canonical prompt-engineer source.

Download SKILL.md Raw portable skill file.

All Skills Browse the full skill catalog.