Written from 15 named sources · Oct 16 · first result A research companion to VO Infusion Episode 6, "Prompt Engineering, Part 1." Researched with Pig Knuckle and reviewed by Deb Munro and Keith Norton, with sources for every claim. Key findings The most reliable prompting habits are explicit context, clear output format, and instructions placed first, with delimiters such as ### or """ to keep direction separate from the script. OpenAI's API guidance is built around that structure. [7] Worked examples (few-shot prompting) can improve format consistency, but they are not a free upgrade. A 2026 study on LLaMA-3 8B Instruct, one run per condition, treated the results as empirical observations: few-shot raised solve rate on structured classification while lowering accuracy and F1 on SST-2, and few-shot accuracy on GSM8K collapsed to 1.7% with a 100% solve rate. [6] Persona prompting can change style, framing, and vocabulary. A 2025 study by Araujo and colleagues, summarized by Emergent Mind, reported mixed, task- and model-dependent effects: up to a 37% strict improvement rate in some settings, declines in 22% of tasks, and drops of up to 30 percentage points from irrelevant persona details. [10] Truong and colleagues (EMNLP 2025) showed that writing style and prompt formatting can change measured benchmark scores even when the meaning of the task stays the same. That is prompt-wording sensitivity, not proof that assigning a role makes a model smarter. [13] Step-by-step reasoning (chain-of-thought) helped solve rate and interpretability on GSM8K in that LLaMA-3 study without beating zero-shot accuracy. As of October 2026, Anthropic's Opus 5.5 guide, reported by Search Engine Journal, suggests chat apps consider removing "think carefully" lines; in Anthropic's own chat test, replies started sooner with no clear decline in quality. Effort is the first API setting to adjust for developers balancing quality, speed, and cost. [1][6] Voice actors already practice the same craft: given circumstances, subtext, status, and what happened two steps before. Those map onto context, constraints, and role. Ask for sources, then verify them yourself. [7][8] For US and Canadian performers, this is booth prep, not legal advice. Union terms, likeness clauses, and licensing belong with SAG-AFTRA, ACTRA, or qualified entertainment counsel. What prompting techniques actually help? Prompt engineering is structuring the text you send a language model so the output is usable. Leave the job unstated and you get the most average completion. State the task, the circumstances, the format, and what to do when something is missing, and you constrain the guess. [7][8] OpenAI's help article is blunt: put instructions at the beginning, separate them from context with ### or """, and be specific about outcome, length, format, and style. Show the desired format with examples when you need structure. Start zero-shot, then add few-shot if that fails. Prefer saying what to do rather than only what not to do. [7] A cross-provider write-up by Steve Kinney describes the same core: separate instructions from data, define failure behavior, and treat XML-style tags as a first-class practice in Claude guidance, with role prompting supported across OpenAI, Anthropic, and Gemini. That piece is secondary commentary, not Google's or Anthropic's own docs, so treat named tags and Gemini "fallback" wording as a summary, not gospel. [8] The 2026 LLaMA-3 paper (Information 17(9), 920, 20 September 2026) used LLaMA-3 8B Instruct, 4-bit quantization, one greedy run per condition, subsets of up to 650 samples, and no significance tests. The authors call the numbers empirical observations, not general laws for frontier models. Few-shot helped format validity (solve rate) on structured classification, not accuracy. On SST-2, few-shot lost to zero-shot on accuracy and F1. On CommonsenseQA, all three methods hit 100% solve rate, with zero-shot slightly ahead on accuracy. On GSM8K, CoT raised solve rate and made the math readable; it did not beat zero-shot accuracy. Few-shot solve rate hit 100% while accuracy fell to 1.7% because the model copied example patterns. [6] As of October 2026, Anthropic's Opus 5.5 prompting guide (accessed 7 October 2026, via Search Engine Journal) says Opus 5.5 decides how much to think. Chat apps should consider removing "think carefully" lines. In Anthropic's chat-product test, removal made replies start sooner with no clear decline in quality. The same reporting notes Anthropic's Opus 4.7 docs still suggest "Think carefully before responding" when effort must stay low. Effort is an API control for developers, not a slider in a consumer chat window. Do not treat removal as a universal rule. [1] Technique What it does Reliability in the sources Booth use Explicit context Domain, audience, job, format [7][8] Strong across vendor guidance Client specs, market, demographic Constraints Length, labels, "unknown" when missing [7][8] Strong when positive and specific No rewrites, no invented claims Few-shot In-context input-output pairs [6][7] Mixed; format up, accuracy can drop [6] Two gold audition notes for shape Chain-of-thought Intermediate steps before the answer [6] Mixed; GSM8K solve rate, not accuracy [6] Dense medical beats, then one verb Role / persona Style and priors [10][13] Variable; style more than "IQ" [10] Casting-director lens, then you choose Does persona prompting improve quality or mainly change style? On the show we said a persona pulls the output off the bland average. Ask a generic bot for audition notes and you get generic notes. Ask it to answer as a seasoned commercial casting director and the vocabulary tightens. That matches how we work: status and character change placement and cadence. They do not raise your IQ. Araujo et al. (27 August 2025), as summarized by Emergent Mind (updated 31 August 2025), evaluated expert personas across 27 tasks and 9 models. Effects were mixed. Some settings showed a 37% strict improvement rate, not a 37% accuracy gain. Performance declined in 22% of tasks, especially smaller models and narrow expert prompts. Irrelevant details (names, hobbies) produced drops of up to 30 percentage points. Constraint language and two-step refinement helped the largest models; smaller models could get worse. Do not treat personas as a general accuracy upgrade. [10] Truong, Fogliato, Heidari, and Wu (EMNLP 2025) rewrote evaluation prompts with persona-based prompting to emulate writing styles. Identical meaning, different style and formatting, different estimated scores. Distinct styles triggered high or low performance across model families. Use that as a warning about brittle wording, not as evidence that "you are a genius" makes the model smarter. [13] The parenting comparison comes from Amanda Askell, Anthropic's in-house philosopher, describing her relationship to Claude during training and character work, as she told Vox. She was talking about shaping Claude's character through the constitution (the "soul doc"), not about a user parenting a finished chatbot in a window. As users we are directing an intern who arrives with a lot of text and no instinct for this particular spot. [15] Treat a persona as a directorial choice. "World-class condiment commercial producer and casting agent" is a lens. Pair it with constraints and a task, then keep the notes you can play. What OpenAI, Anthropic, and Google recommend OpenAI (as of the current Help Center article): use a capable current model; instructions first; ### or """ between instruction and context; specify context, outcome, length, format, and style; show format with examples; start zero-shot, then few-shot, then consider fine-tuning; say what to do instead of only what not to do. [7] Anthropic, as of October 2026 for Claude Opus 5.5 (secondary report on the vendor guide): retest effort rather than reusing Opus 5 settings; medium effort is the default; try lower effort before prompt changes if you need less thinking; chat apps should consider dropping "think carefully" lines; tag pasted outside text. Effort is for API developers. Kinney reports XML tagging as first-class in Claude guidance. [1][8] Google's own prompting pages were not in the supplied research set. Kinney's secondary guide says Gemini guidance converges on clear instructions, explicit constraints, and component placement, and recommends examples, with a warning about too many examples overfitting. Do not invent Google "fallback" wording. [8] Four habits travel: state the job and the standard; isolate instructions from the copy; specify the answer shape before generation; say what to do when the source is thin. [7][8] As of October 2026, use the standard Claude interface. Do not hunt a standalone Cowork product for this work. Translating prompt engineering into the voice actor's toolkit We already parse scripts, hunt subtext, track status, and take direction. Prompting is that craft written down. Susan Blu's Word of Mouth, with Toni Eubanks, is still the classic on digging under the ink. Feed a model copy with no given circumstances and you get mechanical generalities. The tools are like the smartest intern from MIT who knows nothing about your job. Give structure and a launching point. My main prompt runs about eight pages. It started as a paragraph and grew through months of saying what I wanted and did not want. Script analysis: given circumstances and objectives This is explicit context and a stated goal. [7][8] I am auditioning for a 30-second automotive commercial. The speaker is an independent mechanic talking to a car owner whose vehicle has over 150,000 miles. The objective is to reassure without sounding salesy or alarmist. Analyze the copy and list: 1) the speaker's emotional objective, 2) the primary friction, 3) three active verbs for how the speaker attacks the copy. Copy: [Insert script] The moment before: sequential context I ask for what happened two steps before, then improvise my way to the script. That is initial state, not magic. Read this 15-second promo for a sports documentary. Generate two plausible scenarios for what happened two steps before the first word. For each, state the physical action just completed. Copy: [Insert script] Subtext: surface versus meaning Ask for extraction, not a rewrite. CoT-style steps can help on nested dialogue; keep them short. [6] Review the dialogue. In a two-column markdown table, spoken line on the left, unspoken subtext and hesitation on the right. Do not rewrite the dialogue. Dialogue: [Insert dialogue] Status: behavioral constraint Status is a constraint on weight, pace, and authority. [7] Act as a veteran voice director for a commercial banking campaign. The speaker has technical authority and deferential social status toward a small-business founder. Give three performance notes that shift that balance without arrogance or groveling. Copy: [Insert script] Audience and direction notes: boundaries Negative constraints work better when you also say what to do. [7][8] Provide two directorial passes for this public health radio announcement for working parents in Alberta. Rules: Pass 1: grounded, calm, peer-to-peer. Pass 2: urgent, kinetic, brisk. Under 30 words each. Avoid empty adjectives. Use posture, tempo, and placement. Copy: [Insert script] How to ask a model for sources and verify them Unaugmented models predict text. They do not look up a library unless a retrieval tool is attached. Kinney notes brittle prompts can hallucinate citations and recommends tool-grounded retrieval plus checking quotes against source documents. Ask for sources as part of the assignment, then leave the chat to verify. [8] Do not type "give me some sources." Specify document types, required fields, jurisdiction, and what to do when a fact is missing. [7][8] Research non-broadcast digital voiceover usage standards in the United States and Canada. Cite only official agreements, published union rate sheets, or peer-reviewed legal analyses. For every claim: organization, title, year, section. If unconfirmed or jurisdiction-specific, label Unverified or Disputed. Do not invent citations. Then: open the DOI, ISBN, or URL in a browser or registry. Match title and authors. Read the passage. Confirm it supports the exact claim. Check whether US federal or California material applies in British Columbia or Ontario. A trade blog is commentary; the CBA or union notice is primary. Never ask the same model "are you sure?" as proof. This is operational prep, not legal advice. Contracts, likeness, and royalties: SAG-AFTRA, ACTRA, or an entertainment lawyer in your jurisdiction. What this means for your voice business Start a master prompt this month. Mine grew to about eight pages. Draft one page: voice attributes, studio kit, target markets, format rules. Save a text file and paste it in. Write a short "do this instead" list for audition analysis. Ban empty notes and copy rewrites. Add: "Do not rewrite the author's copy. Suggest physical actions, not mood words." Run the next three auditions through it. Separate instructions from copy. Put direction first. Wrap the script in delimiters or tags. That habit alone cuts a lot of confusion. [7][8] A/B persona versus no persona on one real script. Same analysis twice: generic assistant, then veteran automotive casting. Keep playable notes. Discard impressive prose. Zero-trust client research. If you use a model for brand or medical terminology, click every primary link before it leaves your mouth or your email. Keep the last call human. The intern can suggest a circumstance. Your ears, ethics, the client, and the page decide the take. Go further OpenAI, "Best practices for prompt engineering with the OpenAI API." Instruction order, delimiters, examples, and format. [7] "Evaluating Prompt Engineering Techniques for LLaMA-3" (Information, 2026). Zero-shot, few-shot, and CoT on one small model, with solve rate versus accuracy. [6] Araujo et al., "Principled Personas," summarized in Emergent Mind, "Expert Persona Prompting." Mixed gains, 22% declines, irrelevant-detail drops. [10] Truong et al., "Persona-Augmented Benchmarking" (EMNLP 2025). Style and format change measured scores with the same meaning. [13] Search Engine Journal report on Anthropic's Claude Opus 5.5 prompting guide (as of October 2026). Effort first for developers; consider dropping "think carefully" in chat. [1] Steve Kinney, "Prompt Engineering Across the OpenAI, Anthropic, and Gemini APIs." Secondary cross-provider map of structure, roles, and verification. [8] Susan Blu and Toni Eubanks, Word of Mouth. Script analysis and self-direction. Sigal Samuel, Vox, "Claude has an 80-page 'soul document.'" Amanda Askell on character training, not day-to-day user prompting. [15] VO Infusion is a podcast about artificial intelligence and the voice acting industry, co-hosted by voice actor and coach Deb Munro and voice actor Keith Norton. Episodes cover voice cloning and licensing, legal protections for performers, prompt engineering for voice actors, AI tools for the business side of voiceover and why human performance still matters. Listen wherever you get podcasts or on YouTube at @VOInfusion, and find more at voinfusion.com. Pig Knuckle (pigknuckle.ai) is a research orchestration platform from Alchemy Agentic, founded by Paul Langtry and Keith Norton. It breaks a research question into steps, sends each step to the AI model best suited to it, and builds its citation list from the documents it actually read, so every claim can be traced to a source. Sources [1] Anthropic Publishes Prompting Guidance For Claude Opus 5.5 — https://www.searchenginejournal.com/anthropic-claude-opus-5-5-prompting-guidance/591278 [6] Evaluating Prompt Engineering Techniques for LLaMA-3: A Study of Zero-Shot, Few-Shot, and Chain-of-Thought Prompts Across Reasoning and Classification Tasks — https://www.mdpi.com/2078-2489/17/9/920 [7] Best practices for prompt engineering with the OpenAI API OpenAI Help Center — https://help.openai.com/en/articles/6654000-best-practices-for-prompt-engineering-with-the-openai-api [8] Prompt Engineering Across the OpenAI, Anthropic, and Gemini APIs Steve Kinney — https://stevekinney.com/writing/prompt-engineering-frontier-llms [10] Expert Persona Prompting — https://www.emergentmind.com/topics/expert-persona-prompting [13] Persona-Augmented Benchmarking: Evaluating LLMs Across Diverse Writing Styles - ACL Anthology — https://aclanthology.org/2025.emnlp-main.1155 [15] Claude has an 80-page constitution. Is that enough to make it good? Vox — https://www.vox.com/future-perfect/476614/ai-claude-constitution-soul-amanda-askell Argument Collapse: LLMs Flatten Long-Form Public Debate — https://arxiv.org/html/2606.01736v2 Can You Teach an AI Ethics? by Gary Angel Counter Arts Medium — https://medium.com/counterarts/can-you-teach-an-ai-ethics-07d8e244a6c5 Claude Reacts to Amanda Askell. TL;DR — This is a long article, but… by Kevin O'Shaughnessy Medium — https://medium.com/@ZombieCodeKill/claude-reacts-to-amanda-askell-e1904518659f Prompt Engineering by Anthropic Grab the PDF: https ... — https://www.threads.com/@matt_dancho/post/DOOAK-LjqTa/6-prompt-engineering-by-anthropicgrab-the-pdf-httpsdocsanthropiccomendocsbuild-w A guide to OpenAI prompt engineering (2026) eesel AI — https://www.eesel.ai/blog/openai-prompt-engineering Prompt Engineering Guide: System Prompts, Chain-of-Thought, and Structured Outputs (2026) — https://blog.starmorph.com/blog/prompt-engineering-guide-for-gpt-system-prompts Top 10 Free Prompt Engineering Guides 2026: For Beginners and Advanced Users — https://www.amalytix.com/en/blog/free-prompt-engineering-guides Zero-Shot Prompting in 2026: The Practitioner's Guide Adaline — https://www.adaline.ai/blog/what-is-zero-shot-prompting