Optimize LLM Prompts for Quality Outputs

LLM Prompt Optimizer transforms weak, vague, or inconsistent prompts into precision-engineered instructions for any LLM (Claude, Gemini, GPT-4, Llama), applying systematic frameworks from zero-shot to few-shot, chain-of-thought, and structured-output patterns. It starts by diagnosing the weak-prompt problem pattern against a table of five symptoms and fixes: too vague (generic answers, fixed by adding role + context + constraints), no structure (unparseable output, fixed by specifying format explicitly), hallucination (confident wrong answers, fixed by instructing the model to say when unsure), inconsistency (different answers each run, fixed with few-shot examples), and verbosity (fixed with explicit length constraints).

The core method is the RSCIT framework - every optimized prompt states Role (who the AI is in this interaction), Situation (what context it needs), Constraints (rules and limits), Instructions (exactly what to do), and Template (what the output should look like). A worked before/after example turns "Explain machine learning" into a full prompt specifying a senior-ML-engineer role, the audience's actual background, a 200-word limit, no formulas, an analogy requirement, and a prose-only format.

It provides five reusable prompt patterns as copy-ready templates: chain-of-thought (numbered thinking steps before a final answer), few-shot classification (2-3 labeled examples establishing the pattern before the real input), structured JSON output (an explicit schema with raw-JSON-only instructions), hallucination reduction (answer only from provided context, with an exact fallback phrase when the context doesn't contain the answer), and prompt compression (a verbose request rewritten into a shorter one at equivalent quality).

### Compressed (efficient, same quality)
"Analyze this code: explain what it does, how it works, and flag any issues."

Best practices cover specifying output format explicitly, using delimiters to separate instructions from content, testing edge cases (empty input, unusual data), version-controlling system prompts, and preferring positive instructions over negative-only ones ("don't be verbose" without an alternative). A six-item production audit checklist (clear role, defined output format, edge cases handled, appropriate length, tested on 5+ inputs, hallucination risk addressed) and a troubleshooting table cover four common failure modes: the model ignoring format instructions (move format instructions to the end, use "You MUST return only valid JSON"), inconsistent results between runs (lower temperature to 0.0-0.3, add more few-shot examples), prompts that work in a playground but fail in production (verify the system prompt is sent correctly and check token limits), and outputs that run too long (add explicit word/sentence limits).

Use this skill when a prompt returns inconsistent, vague, or hallucinated results, when you need reliable structured/JSON output, when designing system prompts for AI agents or chatbots, when reducing token usage without sacrificing quality, when implementing chain-of-thought reasoning for complex tasks, or when a prompt works on one model but fails on another.

It is a prompt-engineering skill, not a substitute for environment-specific validation, testing, or expert review of the resulting prompt in production - test any optimized prompt on 5+ varied inputs and edge cases before shipping it.

Inputs: a weak, vague, inconsistent, or model-specific-failing prompt, or a task needing structured output, reasoning, or classification.

Outputs: a precision-engineered prompt following the RSCIT framework, using the appropriate pattern (chain-of-thought, few-shot, JSON schema, hallucination-reduction, or compression) for the diagnosed problem.

Applicable to any LLM provider (Claude, Gemini, GPT-4, Llama) since the patterns are prompt-text-level rather than API-specific; complements LLM operations and evaluation skills once a prompt is optimized.

Developers and prompt engineers building AI agents, chatbots, or LLM-powered features who need prompts that produce consistent, structured, low-hallucination output across models.