Prompt Design Recipes & Persona Patterns

Why this playbook matters

Good prompts are the difference between an assistant that helps and one that frustrates. This playbook gives practical recipes and persona patterns you can copy, adapt, and test so teams get consistent, useful outputs faster. It focuses on patterns for repeatable results across summarization, classification, synthesis, extraction, rewriting, and routine business writing.

How to use this guide

Pick a recipe that matches your task. Start with the provided template, adapt the context and examples to your domain, then run the testing checklist. If outputs are unstable, change one variable at a time (temperature, examples, scope, context size) and re-run tests.

Core patterns and templates

1) System / Role prompt pattern

Use a concise system prompt to set high-level behavior and constraints. Example elements: role, expertise level, output format, tone, and guardrails.

2) Few-shot / Example-driven prompts

Few-shot prompts teach structure by example. Provide 2–5 short examples showing input and desired output before asking the model to act on a new input.

3) Persona patterns

Create a persona that includes perspective, knowledge limits, and behavioral rules. Keep persona prompts stable across sessions to maintain consistent voice and decision rules.

4) Context-window and retrieval strategy

Decide what to include in the prompt vs what to retrieve. For high-accuracy tasks, use retrieval-augmented generation: fetch relevant documents, then provide a concise context block (200–800 tokens) with explicit instructions on which parts to use.

Guidelines:

• Keep context focused—only include passages that matter for the task.
• Number or label passages and tell the model which labels are permitted to use.
• When context is long, ask the model to reference the passage label it used for each claim.

5) Temperature & sampling guidance

Temperature (or equivalent) controls creativity vs determinism. Use sampling and top-p settings conservatively for precision tasks.

• Summarization/classification/extraction: low temperature (0–0.3) and deterministic decoding where possible.
• Ideation, drafting, creative synthesis: higher temperature (0.6–1.0) with wider sampling.
• When reliability matters, pair higher temperature with explicit constraints and stronger example-driven prompts.

6) Prompt chaining / orchestration patterns

Complex tasks often work best as a sequence of small, testable calls rather than a single large prompt.

1. Step A: Extract structured facts (low temperature).
2. Step B: Classify or tag facts (low temperature).
3. Step C: Generate synthesis or narrative using the structured facts (controlled temperature, with format constraints).

Chaining improves auditability and makes it easier to rerun a single stage when requirements change.

Practical examples

Summarization (executive brief)

Classification (support tickets)

Extraction (contract clauses)

Testing checklist

1. Define success criteria: format, content elements, accuracy threshold, and response time.
2. Create 10–20 test cases covering normal, edge, and adversarial inputs.
3. Run with low temperature first for structured tasks; compare outputs to expected results.
4. Log failures and categorize: missing info, hallucination, incorrect format, off-tone.
5. Tune examples, context, or chaining strategy; re-test only the altered variable each cycle.
6. When satisfied, run A/B tests on a sample of real inputs before full rollout.

Common mistakes to avoid

• Overloading a single prompt with too many instructions—split into stages.
• Relying on implicit knowledge—prefer explicit examples and format specs.
• No test harness—without tests, reliability degrades in production.
• Failing to define persona constraints—assistants can drift in tone or invent facts.

Next steps and experimentation ideas

Turn these recipes into a "Prompt Lab" where team members can run candidate prompts against sample inputs, capture outputs, and record success metrics. Start with short experiments (10–20 cases) and iterate.

Quick reference

• Precision tasks: low temperature, few-shot examples, explicit format.
• Creative tasks: higher temperature, looser format, multiple generations then selection.
• Complex pipelines: chain extraction → classification → synthesis.
• Always include a fallback: if uncertain, ask a clarifying question or return a 'cannot confidently answer' flag.

Use these recipes as starting points. The most reliable assistants come from disciplined testing, clear persona rules, and incremental refinement.