prompt-engineering

name	prompt-engineering
author	JM Labs (Javier Montaño)
version	2.0.0
description	Design, evaluate, and optimize LLM instruction packages using deterministic pattern selection, source-grounded context, structured output contracts, guardrails, adversarial test cases, and script-backed evaluation packets. Use when the user asks for prompt engineering, system instruction design, few-shot examples, prompt optimization, prompt evaluation, guardrails, meta-prompting, or prompt design. [EXPLICIT]
allowed-tools	["Read","Write","Glob","Grep","Bash"]
sub-agents	["agents/lead.md","agents/support.md","agents/guardian.md","agents/specialist.md"]
knowledge	["knowledge/knowledge-graph.md","knowledge/body-of-knowledge.md"]
templates	["templates/output.html","templates/output.docx.md","templates/output.xlsx.md"]

Prompt Engineering

"A prompt is not a question — it is an architecture for reasoning."

TL;DR

Design, evaluate, and optimize LLM instruction packages. This skill covers the full lifecycle: define task and sources, select the pattern, write the instruction package, create deterministic test cases, validate guardrails, and produce a versioned evaluation packet. [EXPLICIT]

When to Activate

Use this skill when the user asks to:

• design or improve an LLM instruction package
• choose among zero-shot, few-shot, structured-output, system, meta, RAG-grounded, or constitutional/self-critique patterns
• audit a prompt for injection risk, output schema drift, ambiguous instructions, missing examples, or weak guardrails
• create an evaluation matrix for a prompt with test cases and metrics
• adapt a prompt for a target model without inventing model-specific guarantees

Do not use this skill to generate durable agent prompt files; route that work to prompt-creator. Do not use it to create a full agent constitution; route that work to agent-constitution-creator.

Sub-Agents

Agent	Role in Triad	File
prompt-lead	Designs and writes the prompt	agents/lead.md
prompt-support	Reviews for bias, edge cases, injection risk	agents/support.md
prompt-guardian	Evaluates output quality, validates evidence	agents/guardian.md
prompt-specialist	Deep expertise in advanced patterns (meta, constitutional)	agents/specialist.md

Procedure

Step 1: Discover

• Identify the task the prompt must accomplish
• Determine the target model family only from user input or source evidence
• Gather examples of desired input/output pairs
• Capture required sources, forbidden sources, safety boundaries, output schema, and success metrics
• If required source context is missing, return an ask or coverage_gap packet instead of writing a speculative prompt
• Read knowledge/body-of-knowledge.md for pattern catalog
• Check knowledge/knowledge-graph.md for related concepts
• Apply assets/pattern-decision-matrix.json and assets/prompt-engineering-checklist.md

Step 2: Analyze

• Select the pattern:
  • Zero-shot: task is well-defined, model has sufficient training data
  • Few-shot: task needs examples to calibrate output format/style
  • Reasoning scaffold: task requires multi-step reasoning, but hidden reasoning must not be exposed
  • System instruction: task needs persistent behavioral constraints
  • Meta-prompt: task is to generate other prompts
  • Constitutional: task needs value-aligned, self-correcting output
  • Structured output: task needs schema-constrained output
  • RAG-grounded: task must use retrieved context and cite source boundaries
• Evaluate trade-offs: precision vs cost, latency vs quality, generality vs specificity
• Identify guardrails needed (output format, length, safety)
• Record pattern decision, rejected alternatives, and confidence band

Step 3: Execute

• Write the prompt following the selected pattern
• Structure: role → context → task → constraints → output format → examples
• Add guardrails: output schema, safety filters, refusal patterns
• Create at least three deterministic test cases covering happy path, edge case, and adversarial/injection input
• Document the prompt with evidence tags
• Produce a prompt engineering packet with task, target_model, pattern, prompt, guardrails, output_contract, test_cases, metrics, and risks

Step 4: Validate

• Run evaluation suite: accuracy, consistency, edge case handling
• Check for prompt injection vulnerability
• Verify output format compliance
• Validate the packet with scripts/validate_prompt_packet.py
• Generate deliverable using appropriate template only after validation

Deterministic Contract

• Do not invent target-model capabilities, source facts, examples, metrics, dates, or hidden system behavior.
• Use provided examples or clearly marked synthetic fixtures; never imply synthetic fixtures are production evidence.
• Use stable IDs for test cases: PE-001, PE-002, sorted by id.
• Keep hidden reasoning private; ask for concise rationale or decision trace instead of chain-of-thought transcripts.
• Include refusal or escalation behavior for prompt injection, unsupported sources, and schema mismatch.
• If the user wants a reusable prompt file, hand off to prompt-creator after this skill produces the optimization packet.

Quality Criteria

• Pattern selection justified with evidence
• Prompt tested with at least three diverse inputs
• Edge cases identified and handled
• Injection resistance verified
• Output format consistent across runs
• Guardrails prevent harmful/off-topic output
• Evidence tags on all claims
• Prompt engineering packet passes scripts/validate_prompt_packet.py
• assets/prompt-engineering-checklist.md applied

Anti-Patterns

Anti-Pattern	Why It's Bad	Do This Instead
"Just ask nicely"	No structure = inconsistent results	Use role-context-task-format pattern
Massive single instruction	Exceeds attention, dilutes focus	Decompose into focused steps
No examples	Model guesses output format	Add 2-3 few-shot examples
Invented model behavior	Claims unsupported guarantees	State only source-backed model constraints
No evaluation	"It looks right" isn't evidence	Test with diverse inputs, score metrics
Exposed hidden reasoning	Leaks reasoning traces unnecessarily	Ask for concise rationale or decision summary

Related Skills

• ai-safety — Guardrails and output validation
• structured-output — JSON mode, schema-constrained generation
• context-window-management — Token budgeting for long prompts
• rag-patterns — Prompts that integrate retrieved context
• llm-evaluation — Systematic prompt evaluation methods

Knowledge

• knowledge/knowledge-graph.md — Zettelkasten concept map
• knowledge/body-of-knowledge.md — Pattern catalog and references

Templates

• templates/output.html — Branded HTML prompt documentation
• templates/output.docx.md — Word document spec for prompt library
• templates/output.xlsx.md — Evaluation matrix spreadsheet

Assets

• assets/prompt-engineering-checklist.md defines the reusable deterministic review gate.
• assets/pattern-decision-matrix.json defines pattern selection criteria, required evidence, and risk controls.

Scripts

• scripts/validate_prompt_packet.py validates prompt engineering packets.
• scripts/check.sh runs deterministic positive and negative fixtures.

Usage

Example invocations:

• "/prompt-engineering" — Run the full prompt engineering workflow
• "prompt engineering on this project" — Apply to current context

Assumptions & Limits

• Assumes access to project artifacts (code, docs, configs) [EXPLICIT]
• Uses the user's language unless the prompt artifact or target platform requires another language [EXPLICIT]
• Does not replace domain expert judgment for final decisions [EXPLICIT]

Edge Cases

Scenario	Handling
Empty or minimal input	Request clarification before proceeding
Conflicting requirements	Flag conflicts explicitly, propose resolution
Out-of-scope request	Redirect to appropriate skill or escalate