In this guide · 8 sections
Good prompt engineering skills come down to three habits: briefing the AI properly, steering the conversation with targeted follow-ups, and managing what the AI knows about your task. Together they make up Prompting & Communication, one of the five dimensions of AI fluency. This guide shows what strong prompting looks like at every level, gives you six exercises to try on your own work this week, and points to the best free resources to go further.
What Good AI Prompting Skills Look Like
Strong prompting is less about clever phrasing than about giving the AI what a capable colleague would need: the goal, the context, the source material and what a good result looks like. AISA measures it through three skills: prompt design, iterative dialogue, and context and memory management.
Prompt design
Prompt design is how you build the first request. Beginners type one line and accept whatever comes back. Strong prompters state the task, the audience, the constraints and the format, often with an example. At the top end, prompts get shorter rather than longer, because every element is there for a reason the person can explain. Techniques such as few-shot prompting and negative constraints are tools for that, not goals in themselves.
Iterative dialogue
Iterative dialogue is what you do after the first answer. "Make it better" rarely helps; naming the specific problem does. Skilled practitioners plan multi-turn sequences, use diagnostic follow-ups to find out why an answer went wrong, and use iteration to explore options rather than only to polish one.
Context and memory management
Context management is deciding what the AI should know, and when to start over. It includes chunking long inputs, resetting a conversation that has drifted, and building reusable instructions through system prompts, projects and memory. This is the skill people now call context engineering.
Prompt Engineering Skills at Every Level
AISA scores Prompting & Communication from 1 to 10, on the same scale as the rest of the AISA rubric. This is what each band looks like.
| Level | Score | What it looks like |
|---|---|---|
| Novice | 1–2 | One-line requests. Takes the first answer and rarely follows up. No sense of what the AI already knows. |
| Developing | 3–4 | Uses a technique such as a role or an example now and then. Follow-ups are generic ("try again"). Pastes context when they remember to. |
| Competent | 5–6 | Every prompt has a task, context and format. Follow-ups name the specific issue. Chunks long inputs and resets a conversation on purpose. |
| Proficient | 7–8 | Adapts the structure to the task and the model. Plans multi-turn sequences. Keeps reusable instructions and tells the AI how to work with them. |
| Expert | 9–10 | Prompting is a way of thinking: constraints and meta-instructions from first principles, iteration used to explore, reusable context systems built for others. |
Across 1,800 assessments, the averages are 5.2 for Prompt Design, 5.0 for Iterative Dialogue and 5.3 for Context & Memory Management. The average professional writes a reasonable first prompt; the jump to Proficient comes from what happens after it.
Which Prompting Techniques Actually Work
Some popular prompting advice holds up and some does not. Research from Wharton's Generative AI Labs is a useful check before you build habits around a trick.
Personas change tone, not accuracy
Opening with "You are a world-class expert" feels powerful. In Prompting Science Report 4, expert personas matched to the question's field had no significant effect on factual accuracy for five of the six models tested, and low-knowledge personas often reduced it. Stating explicit criteria for a good answer is the better use of those words.
"Think step by step" matters less with reasoning models
In Prompting Science Report 2, asking reasoning models to think step by step changed average accuracy by about three points either way while making answers slower. Models that already reason do not need the instruction; your effort is better spent on context.
Iteration is the skill most people are dropping
Anthropic's Economic Index found that the share of "directive" Claude conversations, where the user gives a task and Claude completes it with minimal back-and-forth, rose from 27% to 39% between late 2024 and August 2025. Delegating more is not a problem in itself; losing the habit of steering is.
The Best Free Resources to Improve Your Prompting
These resources come from the teams that build the models and from practitioners who test them. All are free.
Start here
Foundational · 4- GuideFreeGetting started with AI: Good enough prompting Ethan MollickTreat the AI as a coworker: give context every time, ask for options, work in dialogue
- GuideFreeGet started writing prompts in Microsoft Copilot MicrosoftA four-part prompt structure: goal, context, source and expectations
- DocsFreeWhat are projects? Claude Help CenterReusable workspaces with their own instructions, files and history
- DocsFreeUsing chat search and memory Claude Help CenterHow AI memory is stored and how to view, edit, pause or reset it
Build on it
Intermediate · 4- VideoFreeAI prompt engineering: A deep dive AnthropicFour prompt specialists on how they write, test and iterate prompts
- GuideFreeBest practices for prompt engineering AnthropicCore and advanced techniques, plus fixes for generic or made-up answers
- InteractiveFreePrompt Engineering Interactive Tutorial AnthropicNine hands-on chapters with exercises, from structure to avoiding hallucinations
- DocsFreePrompt design strategies GoogleClear instructions, examples, prompt chains and specific iteration strategies
For a reference prompt sequence from a top-scoring assessment, read What a 9 out of 10 Looks Like.
How to Improve Your Prompting Skills: 6 Exercises
Each exercise takes 10 to 30 minutes and uses a real task from your week. The goal is to notice what changes the output, so you build judgement rather than a collection of templates.
- Exercise 115 min
Rebuild one real prompt
Take a request you sent an AI this week and rewrite it with four parts: goal, context (who it is for and why), source material, and expected format and length. Run the old and new versions in two fresh chats. Notice which part changed the output most, and what the first version left the AI to guess.
- Exercise 215 min
Let the model interview you
Start a real task with: "Before you start, ask me up to five questions you need answered to do this well." Answer them, then let it work. The questions that surface context you would not have offered are the ones to include by default.
- Exercise 315 min
One change per follow-up
Generate a first draft, then send three follow-ups that each change exactly one thing: audience, length, structure. In a separate chat, send the same first prompt followed by a single "make it better". Notice how targeted follow-ups converge while vague ones drift.
- Exercise 420 min
Examples versus description
For a format you produce often, such as a status update, run the task once with only a written description and once with one or two anonymised examples of good past versions. Notice how much closer the example-led version gets, and whether it copied content it should not have.
- Exercise 515 min
Long-chat reset
In a long conversation, ask the AI to list every requirement and decision agreed so far and check it against your own list. Paste the corrected summary into a fresh chat and rerun the latest task. Notice what had been dropped and whether the fresh chat does better.
- Exercise 625 min
Build one reusable setup
Create a project or custom instructions for one recurring piece of work: your role, audience, house style and things to avoid. Test it on three different tasks, then review what the assistant has stored in memory. Notice what belongs in standing instructions and what belongs in each prompt.
Curious about your AI Fluency?
AISA helps you measure, prove and improve your AI skills — free report in a 20–40 minute chat.
AISApedia Concepts to Learn Next
These are the AISApedia entries behind Prompting & Communication, grouped from first steps to advanced. Each explains one technique in plain language and links a hand-picked resource to go further.
Foundational
Start here5 concepts- Conversation Planning
Conversation planning is the practice of structuring AI interactions as a deliberate sequence of information disclosure — context first, then constraints, then exploration, then implementation — rather than dumping requirements in a single message.
ArticleGetting started with AI: Good enough prompting - Iterative Refinement
Iterative refinement is the practice of improving AI outputs through successive rounds of targeted feedback rather than attempting to get a perfect result from a single prompt.
DocsWhat is Iterative Prompting? - Output Formatting
Output formatting is the practice of controlling the structure, style, and presentation of AI-generated responses through concrete examples rather than abstract instructions.
DocsPrompt Engineering Tips: Specificity, Format Examples, and Delimiters - Role Prompting
Role prompting is the technique of assigning a specific persona, profession, or expertise profile to a language model at the start of a conversation.
TutorialRole Prompting: Guide LLMs with Persona-Based Tasks - System Prompts
A system prompt is an instruction block provided to a language model before any user messages, establishing the model's behaviour, persona, constraints, and domain context for the entire conversation.
DocsGiving Claude a Role with a System Prompt
Intermediate
Build on it14 concepts- A/B Prompt Testing
A/B prompt testing is the practice of systematically comparing two or more prompt variants against the same input to measure which produces better results according to defined evaluation criteria.
ArticleA/B testing for LLM prompts: A practical guide - Chain-of-Thought Prompting
Chain-of-thought prompting is a technique that instructs AI models to break down complex problems into explicit intermediate reasoning steps before producing a final answer.
ArticleChain-of-Thought Prompting - Learn Prompting - Context Compression
Context compression is the practice of reducing the length of input text provided to an AI model while preserving informational density, so that the model's finite context window is filled with high-signal content rather than redundant phrasing.
TutorialHow to Compress Your Prompts and Reduce LLM Costs - Conversation Branching
Conversation branching is an AI workflow design pattern where the initial interaction diagnoses user intent and routes the conversation to a specialised handling path, rather than following a single linear sequence.
ArticleBranching Conversations in ChatGPT Explained: How It Works! - Conversation Chunking
Conversation chunking is the practice of deliberately ending an AI conversation and starting a new one with a synthesised summary of prior context, rather than continuing a single thread indefinitely.
ArticleContext Management: Handling Long AI Conversations & Documents - Diagnostic Follow-Ups
Diagnostic follow-ups are targeted questions asked after an AI's initial response to probe for hidden assumptions, edge cases, failure modes, and alternative approaches that the first answer omitted.
ArticlePrompt Debugging: Diagnosing and Fixing Broken LLM Outputs for Reliable AI Performance - Few-Shot Prompting
Few-shot prompting is a technique where two to five examples of desired input-output pairs are included in the prompt before the actual task, enabling the model to infer the expected pattern, format, tone, and quality level from the examples rather than from explicit instructions alone.
DocsFew-Shot Prompting - Image Generation Prompting
Image generation prompting is the practice of constructing structured, layered prompts for AI image generators (DALL-E, Midjourney, Stable Diffusion, Ideogram) that specify subject, artistic style, and technical parameters separately.
ArticleHow to write AI image prompts like a pro - Meta-Prompting
Meta-prompting is the practice of asking an AI model to generate, refine, or optimise prompts for itself or for other models.
ArticleA Complete Guide to Meta Prompting - Multi-Modal Prompting
Multi-modal prompting is the technique of combining text, images, audio, or other media types in a single AI prompt to provide richer context than any single modality alone.
DocsVision - Claude API Docs - Negative Constraints
Negative constraints are prompt instructions that specify what the AI model should avoid rather than what it should produce.
ArticleNegative Prompting and How to Tell AI What NOT to Do - Prompt Debugging
Prompt debugging is the systematic process of diagnosing why an AI model's output does not match expectations.
ArticleSystematic Prompt Engineering: From Trial and Error to Data-Driven Optimization - Prompt Templates
Prompt templates are parameterised prompt structures with fixed instructional scaffolding and variable input slots, designed for repeated use across different inputs and contexts.
ArticlePrompt Patterns: What They Are and 16 You Should Know - Structured Output Formats
Structured output formats are explicit schemas — JSON, CSV, XML, markdown tables, or other machine-readable formats — specified in AI prompts to constrain the model's response into a predictable, parseable structure.
ArticleThe Guide to Structured Outputs and Function Calling with LLMs
Advanced
Go deep3 concepts- Context Engineering
Context engineering is the discipline of designing what information an AI system receives, in what structure, and at what granularity — optimising for relevance within the constraints of a finite context window.
ArticleEffective Context Engineering for AI Agents - Cross-Session Context
Cross-session context refers to the ability of an AI system to retain and reference information from previous conversations, enabling continuity across interactions that would otherwise be stateless.
ArticleMemory in Agents: What, Why and How - Structured Output Parsing
Structured output parsing is the practice of instructing language models to return responses in machine-readable formats — JSON, XML, or typed schemas — rather than free-form text.
DocsStructured Model Outputs
Measure Your Prompting Skills
Prompting is easy to overrate in yourself. In the AISA assessment you talk with Aisa, an AI interviewer, about how you really brief, steer and manage AI, and your report scores Prompting & Communication alongside the other four dimensions of AI fluency, with the specific techniques to work on next. Take the free AI fluency assessment to see where you stand.
Related reading: Prompting & Communication: Score Bands Explained — what each score band looks like in real answers. Related reading: How Good Am I at Prompting? — a quick self-test before you take the full assessment. Related reading: Critical Thinking with AI — how to check what your prompts produce.
Frequently Asked Questions
How can I improve my AI prompting skills quickly?
Rewrite one real prompt with four parts (goal, context, source material, expected format) and compare it with the original in a fresh chat. Then practise targeted follow-ups that change one thing at a time. A week of this on your own work does more than any list of prompt templates.
Is prompt engineering still a useful skill?
Yes, though it has shifted. Newer models need fewer tricks, so the value has moved from special phrasing to clear briefing, targeted iteration and managing context. Those habits are what separate strong AI users from average ones.
Do I need to tell the AI to act as an expert?
Not for accuracy. Wharton research found that expert personas did not reliably improve factual accuracy, although they can change tone. You get more from stating what a good answer must include and what to do when the AI is unsure.
What is the difference between prompting and context engineering?
Prompting is how you phrase a single request. Context engineering is deciding everything the AI sees: instructions, documents, memory and conversation history. As tasks get longer and more automated, context engineering matters more than any single prompt.
Curious about your AI Fluency?
AISA helps you measure, prove and improve your AI skills — free report in a 20–40 minute chat.
