Prompt Engineering

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

TL;DR

Design, evaluate, and optimize prompts for any LLM. This skill covers the full lifecycle: understanding the task → selecting the right pattern (few-shot, CoT, system prompt, meta-prompt) → writing the prompt → evaluating output quality → iterating. Produces Zettelkasten-ready knowledge artifacts and branded deliverables (HTML, DOCX, XLSX). [EXPLICIT]

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 LLM (Claude, Gemini, GPT, Llama — affects syntax)
• Gather examples of desired input/output pairs
• Read knowledge/body-of-knowledge.md for pattern catalog
• Check knowledge/knowledge-graph.md for related concepts

Step 2: Analyze

• Select the prompt pattern:
  • Zero-shot: task is well-defined, model has sufficient training data
  • Few-shot: task needs examples to calibrate output format/style
  • Chain-of-thought (CoT): task requires multi-step reasoning
  • System prompt: task needs persistent behavioral constraints
  • Meta-prompt: task is to generate other prompts
  • Constitutional: task needs value-aligned, self-correcting output
• Evaluate trade-offs: precision vs cost, latency vs quality, generality vs specificity
• Identify guardrails needed (output format, length, safety)

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
• Test with 3+ diverse inputs to verify robustness
• Document the prompt with evidence tags

Step 4: Validate

• Run evaluation suite: accuracy, consistency, edge case handling
• Check for prompt injection vulnerability
• Verify output format compliance
• Score confidence (>= 0.95 per Constitution)
• Generate deliverable using appropriate template

Quality Criteria

• Pattern selection justified with evidence
• Prompt tested with 3+ 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
• Knowledge graph updated with new patterns discovered

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 prompt	Exceeds attention, dilutes focus	Decompose into chain of focused prompts
No examples	Model guesses output format	Add 2-3 few-shot examples
Ignoring the model	Claude ≠ GPT ≠ Llama	Adapt syntax to target model
No evaluation	"It looks right" isn't evidence	Test with diverse inputs, score metrics

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

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]
• Requires English-language output unless otherwise specified [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