# Excellent AI Prompts: full reference text > The AI Encyclopedia and the AI statistics from https://www.excellentaiprompts.com, by Lea Thomas. The short guide to the site is at https://www.excellentaiprompts.com/llms.txt. --- # AI statistics ## AI adoption statistics These figures show how far AI use has spread among organizations, employees and the general public. Each figure shows its source, the period it covers and what the source measured. ### Key takeaways - 88% of organizations use AI, according to the Stanford 2026 AI Index. - 44% of U.S. adults said they use ChatGPT in February 2026, up from 18% in 2023, according to Pew Research Center. - 52% of U.S. employees used AI at work at least a few times a year in May 2026, and 15% used it daily, according to Gallup. ### Organizations - **88%** of organizations use AI, the 2026 AI Index found. The report describes this figure as organizational adoption. What this means: AI is now normal in most organizations. Company use does not mean that each person in the company uses AI well, or uses it at all. Period: 2026 report. Source: [Stanford HAI, The 2026 AI Index Report](https://hai.stanford.edu/ai-index/2026-ai-index-report). - **$285.9 billion** of private money was invested in AI in the United States in 2025. The same report puts the figure for China at $12.4 billion. What this means: Money keeps flowing into AI in the United States. Expect new products and frequent changes to the tools you already use. Period: 2025. Source: [Stanford HAI, The 2026 AI Index Report](https://hai.stanford.edu/ai-index/2026-ai-index-report). ### The general public - **44%** of U.S. adults said they use ChatGPT in February 2026. The figure was 18% in 2023, 23% in 2024 and 34% in 2025. Pew changed the wording of the question in 2026, so the years are not a perfect match. What this means: Many of your customers and colleagues have now tried an AI chat tool. Basic familiarity is common, so the advantage comes from what you do after the first chat. Period: February 17 to 23, 2026. Source: [Pew Research Center, Americans and AI 2026: Chatbots, Smart Devices and Views on Impact](https://www.pewresearch.org/chart/the-share-of-americans-who-say-they-use-chatgpt-has-more-than-doubled-since-2023/). - **53%** population adoption of generative AI, three years after its release. The report ranks the United States 24th among countries, at 28.3%. Its measure differs from the Pew survey above. What this means: Generative AI spread to half of the population in three years. The United States is behind many countries on this measure, so daily AI use still sets a U.S. professional apart. Period: 2026 report. Source: [Stanford HAI, The 2026 AI Index Report](https://hai.stanford.edu/ai-index/2026-ai-index-report). - **Nearly 40%** of the U.S. population aged 18 to 64 used generative AI in August 2024, at work or at home. The figure comes from a nationally representative survey by the economists Bick, Blandin and Deming. What this means: AI use was already common in 2024, at work and at home. Compare this figure with the newer ones on this page to see how fast use grows. Period: August 2024. Source: [Federal Reserve Bank of St. Louis, Generative AI, Productivity and the Future of Work](https://www.stlouisfed.org/open-vault/2025/oct/generative-ai-productivity-future-work). ### Employees - **52%** of U.S. employees used AI in their role a few times a year or more. This is the widest of the three Gallup measures. What this means: Half of employees have tried AI at work. A few uses a year is not the same as a work habit. Period: May 2026. Source: [Gallup, Indicator: Artificial Intelligence](https://www.gallup.com/699797/indicator-artificial-intelligence.aspx). - **30%** of U.S. employees used AI in their role a few times a week or more. This group uses AI as a regular part of the work week. What this means: About one employee in three uses AI each week. This group gets regular practice, and practice teaches people which tasks suit AI. Period: May 2026. Source: [Gallup, Indicator: Artificial Intelligence](https://www.gallup.com/699797/indicator-artificial-intelligence.aspx). - **15%** of U.S. employees used AI in their role daily. Daily use is still a minority habit. What this means: Daily users are a small group. A daily AI habit still puts a professional ahead of most colleagues. Period: May 2026. Source: [Gallup, Indicator: Artificial Intelligence](https://www.gallup.com/699797/indicator-artificial-intelligence.aspx). Page: https://www.excellentaiprompts.com/statistics/ai-adoption-statistics (figures checked 2026-11-10) ## AI productivity statistics These figures come from controlled experiments and a national survey. They show what generative AI changed when professionals used it for writing, consulting and customer support work. ### Key takeaways - Generative AI users reported saving 5.4% of their work hours, about 2.2 hours in a 40-hour week, according to the St. Louis Fed. - In a controlled experiment published in Science, ChatGPT cut the time for professional writing tasks by 40% and raised quality by 18%. - The gains are larger for less experienced workers, and AI lowers accuracy on tasks it is not suited to. ### Time saved - **5.4%** of work hours were saved, on average, by workers who used generative AI. That is roughly 2.2 hours in a 40-hour work week. The figure comes from what users reported in a survey. What this means: The average saving is modest: about two hours in a 40-hour week. Your own saving depends on how much of your work suits AI. Period: Published October 2025. Source: [Federal Reserve Bank of St. Louis, Generative AI, Productivity and the Future of Work](https://www.stlouisfed.org/open-vault/2025/oct/generative-ai-productivity-future-work). - **40%** less time was needed for professional writing tasks when people used ChatGPT. The experiment had 453 college-educated professionals. Half of them, chosen at random, were given ChatGPT. What this means: Writing is one of the tasks where AI saves the most time. Start with drafts, summaries and emails if you want fast results. Period: 2023. Source: [Science, Experimental evidence on the productivity effects of generative artificial intelligence (Noy and Zhang)](https://doi.org/10.1126/science.adh2586). - **25.1%** faster task completion for consultants who used AI on tasks suited to it. The experiment had 758 consultants at Boston Consulting Group. They also completed 12.2% more tasks. What this means: On the right tasks, AI makes skilled professionals faster. The condition is that the task suits AI. Period: 2023. Source: [Harvard Business School working paper, Navigating the Jagged Technological Frontier (Dell'Acqua and others)](https://papers.ssrn.com/abstract=4573321). ### Quality and output - **18%** higher output quality in the same writing experiment. The time and quality figures come from the same 453 professionals. What this means: Speed did not cost quality in this study. The professionals worked faster and produced better work at the same time. Period: 2023. Source: [Science, Experimental evidence on the productivity effects of generative artificial intelligence (Noy and Zhang)](https://doi.org/10.1126/science.adh2586). - **15%** more customer issues resolved per hour, on average, by support agents with an AI assistant. The study followed 5,172 customer support agents. An earlier working paper version reported 14%. Less experienced agents gained the most. What this means: AI help gives the largest gains to people who are new to a job. A team can use AI to bring new staff up to speed. Period: Published 2025. Source: [The Quarterly Journal of Economics, Generative AI at Work (Brynjolfsson, Li and Raymond)](https://doi.org/10.1093/qje/qjae044). ### Where AI made work worse - **19 percentage points** lower chance of a correct solution when consultants used AI on a task outside its abilities. The authors call the uneven edge of AI ability the jagged frontier. The same consultants gained on tasks inside it. What this means: AI can make results worse on the wrong task. Decide whether a task suits AI before you trust the answer. Period: 2023. Source: [Harvard Business School working paper, Navigating the Jagged Technological Frontier (Dell'Acqua and others)](https://papers.ssrn.com/abstract=4573321). Page: https://www.excellentaiprompts.com/statistics/ai-productivity-statistics (figures checked 2026-11-10) ## AI reliability and risk statistics These figures show the limits of current AI systems. They are the reason a professional checks AI work before it goes to a client. ### Key takeaways - AI agents completed 66.3% of computer tasks on the OSWorld test in 2025. The human baseline is 72.35%, according to the Stanford 2026 AI Index. - The top model read an analog clock correctly 50.1% of the time in the same report. - Documented AI incidents rose to 362 in 2025, from 233 in 2024. ### How often AI gets the task right - **66.3%** of computer tasks on the OSWorld test were completed by AI agents in 2025. The human baseline on the same test is 72.35%. OSWorld tests an agent on tasks in real computer programs. What this means: AI agents now complete most tasks on this test, but they still fail about one task in three. Review agent work before you rely on it. Period: 2025. Source: [Stanford HAI, The 2026 AI Index Report](https://hai.stanford.edu/ai-index/2026-ai-index-report). - **50.1%** of the time, the top AI model read an analog clock correctly. The report uses this as an example of uneven ability. The same models pass very hard tests in other areas. What this means: A model that passes hard tests can still fail a simple one. Do not assume that skill in one area carries over to another. Period: 2026 report. Source: [Stanford HAI, The 2026 AI Index Report](https://hai.stanford.edu/ai-index/2026-ai-index-report). - **19 percentage points** lower chance of a correct solution when consultants used AI on a task outside its abilities. The experiment had 758 consultants at Boston Consulting Group. What this means: Trust in AI can lower the quality of work when the task is outside AI ability. Human judgment is still necessary. Period: 2023. Source: [Harvard Business School working paper, Navigating the Jagged Technological Frontier (Dell'Acqua and others)](https://papers.ssrn.com/abstract=4573321). ### Incidents and trust - **362** AI incidents were documented in 2025, up from 233 in 2024. The report tracks publicly documented incidents year by year. What this means: Documented AI problems increase as use increases. A check step in your workflow protects you and your customers. Period: 2025. Source: [Stanford HAI, The 2026 AI Index Report](https://hai.stanford.edu/ai-index/2026-ai-index-report). - **31%** of people in the United States trust their government to regulate AI. This is the lowest figure among the countries in the report. Singapore is the highest, at 91%. What this means: Most people in the United States do not trust the government to manage AI risk. Clear rules about how your business uses AI can build trust with your customers. Period: 2026 report. Source: [Stanford HAI, The 2026 AI Index Report](https://hai.stanford.edu/ai-index/2026-ai-index-report). Page: https://www.excellentaiprompts.com/statistics/ai-reliability-statistics (figures checked 2026-11-10) --- # AI Encyclopedia ## What is an agent skill? An agent skill is a saved package of instructions, and sometimes files and scripts, that teaches an AI agent how to do one specific kind of task your way. ### Key takeaways - A skill is reusable. You write it one time and the agent uses it when the task comes up. - It usually contains a short description, step-by-step instructions and examples. - Skills let you build your methods into the tool, so you do not explain them in every chat. ### How it works You write down how a task is done, with the steps, the rules and a sample of good output, and save it in the format your AI tool uses. The agent sees a short description of each skill. When a request matches, it loads the full instructions and follows them. ### Example A consultant saves a skill called Client Report. It contains the report structure, the firm's tone rules and a finished sample. When the consultant says "write the monthly report for this client", the agent loads the skill and produces a report in the house format. ### Why it matters Skills turn your know-how into something an agent can run again and again. They are how a solo operator gets consistent work from AI with no long prompt each time. ### Common mistake A skill from an unknown source can contain instructions you did not expect. Read a skill before you install it, in the same way you check software before you run it. ### Related terms - [AI Agent](https://www.excellentaiprompts.com/ai-encyclopedia/ai-agent) - [System Prompt](https://www.excellentaiprompts.com/ai-encyclopedia/system-prompt) - [Agentic Workflow](https://www.excellentaiprompts.com/ai-encyclopedia/agentic-workflow) - [Model Context Protocol (MCP)](https://www.excellentaiprompts.com/ai-encyclopedia/model-context-protocol) Source: https://www.excellentaiprompts.com/ai-encyclopedia/agent-skill (Excellent AI Prompts, updated 2026-11-10) ## What is an agentic workflow? An agentic workflow is a business process in which one or more AI agents do a sequence of steps, with fixed checkpoints and limits set by a person. ### Key takeaways - The workflow sets the route. The agent makes the judgment calls on the way. - It is more predictable than a fully free agent and more flexible than classic automation. - Good workflows include a check step before anything is sent or published. ### How it works You define the stages of a process, such as collect, draft, check and deliver. An agent handles each stage with its own instructions and tools. Between stages, a rule or a person checks the output. The process runs the same way each time, but the agent adapts to each new input. ### Example A newsletter writer builds a workflow. One agent gathers the week's sources. A second drafts the issue in the writer's voice. A third checks each claim against the sources. The writer reads the draft and the check report, then publishes. ### Why it matters Most business value from agents comes from repeatable workflows, not from one large open task. A workflow is also easier to test and to improve. ### Common mistake People automate a process that does not yet work by hand. Run the steps yourself first, write down what good output looks like, and then delegate the work. ### Related terms - [AI Agent](https://www.excellentaiprompts.com/ai-encyclopedia/ai-agent) - [AI Loop](https://www.excellentaiprompts.com/ai-encyclopedia/ai-loop) - [Human in the Loop](https://www.excellentaiprompts.com/ai-encyclopedia/human-in-the-loop) - [Agent Skill](https://www.excellentaiprompts.com/ai-encyclopedia/agent-skill) Source: https://www.excellentaiprompts.com/ai-encyclopedia/agentic-workflow (Excellent AI Prompts, updated 2026-11-10) ## What is an AI agent? An AI agent is an AI system that works toward a goal across many steps. It decides what to do next, uses tools such as search, files and apps, and checks its own progress. ### Key takeaways - A chatbot answers one message. An agent completes a task. - Agents use tools. They can search the web, read and write files, and run code. - The more an agent can do alone, the more its limits and permissions matter. ### How it works You give the agent a goal. The model plans a first step, calls a tool, reads the result and decides the next step. It repeats this cycle until the goal is met or it needs your input. The model supplies the judgment, and the tools let it act. ### Example You ask an agent to find ten podcasts that interview consultants, check that each one took a guest in the last 90 days, and draft a pitch email for each. The agent searches, opens each site, checks the dates and returns a table with ten drafts. ### Why it matters Agents move AI from writing help to delegated work. For a solo business, an agent can do research, follow-up and reporting that used to need an assistant. ### Common mistake An agent can repeat an early error across every later step. Start with small tasks, limit what it may change or send, and review the result before it goes out. ### Related terms - [Agentic Workflow](https://www.excellentaiprompts.com/ai-encyclopedia/agentic-workflow) - [AI Loop](https://www.excellentaiprompts.com/ai-encyclopedia/ai-loop) - [Model Context Protocol (MCP)](https://www.excellentaiprompts.com/ai-encyclopedia/model-context-protocol) - [Human in the Loop](https://www.excellentaiprompts.com/ai-encyclopedia/human-in-the-loop) Source: https://www.excellentaiprompts.com/ai-encyclopedia/ai-agent (Excellent AI Prompts, updated 2026-11-10) ## What is an AI loop? An AI loop is a setup in which an AI agent repeats a cycle of work, check and improve until the result passes a test you defined. ### Key takeaways - A loop needs a clear score or test, so the agent knows when it is done. - The agent checks its own work each round, so quality rises with no extra effort from you. - A stop rule, such as a maximum number of rounds, keeps the loop from running forever. ### How it works You give the agent a goal, a way to measure the result and a stop rule. The agent produces a version, measures it, finds the weakest part and fixes it. It repeats the cycle until the score is high enough or the stop rule applies. Then it reports what it did. ### Example You ask an agent to improve a landing page until ten test readers, played by the model, can each state the offer in one sentence. The agent rewrites, tests and repeats, and it stops after the test passes or after eight rounds. ### Why it matters One prompt gives one draft. A loop gives a draft that already went through several rounds of review. It is the simplest way to get better work from the same model. ### Common mistake A loop improves only what its test measures. If the test is weak, the agent gets a high score on weak work. Write the test with care. ### Related terms - [AI Agent](https://www.excellentaiprompts.com/ai-encyclopedia/ai-agent) - [Agentic Workflow](https://www.excellentaiprompts.com/ai-encyclopedia/agentic-workflow) - [Evaluation (Evals)](https://www.excellentaiprompts.com/ai-encyclopedia/evaluation) Source: https://www.excellentaiprompts.com/ai-encyclopedia/ai-loop (Excellent AI Prompts, updated 2026-11-10) ## What is artificial intelligence (AI)? Artificial intelligence (AI) is software that does tasks that normally need human thinking, such as reading, writing, planning and recognizing images. ### Key takeaways - AI is a broad field. Chat tools such as ChatGPT, Claude and Gemini are one part of it. - Most AI you use at work today is generative AI built on large language models. - AI predicts a likely answer. It does not know facts the way a person does, so you check its work. ### How it works Older software follows rules that a programmer wrote by hand. AI systems learn from examples. A developer shows the system a very large amount of data, and the system adjusts millions or billions of internal numbers until its output matches the examples. After that training, the system can handle new input it has never seen. ### Example A consultant pastes a 40-page client report into an AI chat tool and asks for the five decisions the client must make this quarter. The tool reads the report and returns a short list in seconds. ### Why it matters AI is now a general work tool, like a spreadsheet. People who know what it does well, and where it fails, get more done with the same hours. ### Common mistake People treat AI output as a finished answer. Treat it as a strong first draft from a fast assistant that can be wrong. ### Related terms - [Machine Learning (ML)](https://www.excellentaiprompts.com/ai-encyclopedia/machine-learning) - [Generative AI](https://www.excellentaiprompts.com/ai-encyclopedia/generative-ai) - [Large Language Model (LLM)](https://www.excellentaiprompts.com/ai-encyclopedia/large-language-model) Source: https://www.excellentaiprompts.com/ai-encyclopedia/artificial-intelligence (Excellent AI Prompts, updated 2026-11-10) ## What is a context window? A context window is the maximum amount of text, measured in tokens, that an AI model can read and consider at one time. ### Key takeaways - The context window covers your instructions, your files, the conversation so far and the answer. - When a conversation grows past the window, the tool drops or shortens the oldest parts. - A model uses only what is in the window. It does not remember your other chats unless the tool adds them. ### How it works Each time you send a message, the tool sends the model everything it needs in one block. That block contains the system prompt, the earlier messages, any files and your new message. The block must fit in the context window. Text that does not fit is cut or summarized first. ### Example You work with an AI tool on a long proposal for two hours. Late in the session it forgets a rule you gave at the start. The early messages no longer fit in the window, so the model cannot see them. ### Why it matters Most cases of an AI that forgot something are context window cases. A short reminder of the key rules, or a fresh chat with a clear brief, fixes them. ### Common mistake A large window is not the same as good attention. Models can miss details in the middle of very long input, so put the most important instructions at the start or the end. ### Related terms - [Token](https://www.excellentaiprompts.com/ai-encyclopedia/token) - [System Prompt](https://www.excellentaiprompts.com/ai-encyclopedia/system-prompt) - [Retrieval-Augmented Generation (RAG)](https://www.excellentaiprompts.com/ai-encyclopedia/retrieval-augmented-generation) Source: https://www.excellentaiprompts.com/ai-encyclopedia/context-window (Excellent AI Prompts, updated 2026-11-10) ## What is an embedding? An embedding is a list of numbers that represents the meaning of a piece of text, so that software can compare texts by meaning and not only by exact words. ### Key takeaways - Texts with similar meaning get similar numbers. - Embeddings power search by meaning, recommendations and RAG. - You rarely see them, but they sit behind most document chat features. ### How it works A model reads a passage and outputs a long list of numbers. Passages about the same idea land close together in that number space, even when they share no words. To search, the tool turns your question into an embedding and finds the stored passages that sit closest to it. ### Example You search a help center for "money back". A keyword search finds nothing, because the article says "refund policy". A search built on embeddings finds the article, because the meanings match. ### Why it matters When a product claims semantic search or says it understands your documents, embeddings do that work. They also explain why such a search sometimes returns a related but wrong passage. ### Common mistake Close in meaning is not the same as correct. An embedding search returns the nearest text, and that text can miss the point of your question. ### Related terms - [Retrieval-Augmented Generation (RAG)](https://www.excellentaiprompts.com/ai-encyclopedia/retrieval-augmented-generation) - [Large Language Model (LLM)](https://www.excellentaiprompts.com/ai-encyclopedia/large-language-model) Source: https://www.excellentaiprompts.com/ai-encyclopedia/embedding (Excellent AI Prompts, updated 2026-11-10) ## What is an AI evaluation (eval)? An evaluation, or eval, is a repeatable test that measures how well an AI model, prompt or agent does a specific task. ### Key takeaways - An eval is a set of test cases with a way to score each result. - It tells you if a change to a prompt or a model made things better or worse. - A small eval of 10 to 20 real cases beats a general feeling that the output looks fine. ### How it works You collect real examples of the task, each with a known good answer or a clear standard. You run the AI on all of them and score the results, by hand, with a rule or with a second model that grades. You run the same set again each time you change something and compare the scores. ### Example Before a consultant switches the model behind a proposal prompt, they run both models on 12 past client briefs and grade each draft against a five-point checklist. The new model wins on nine briefs, so they switch. ### Why it matters Models and prompts change often. An eval lets you decide with evidence, and it catches a quality drop before your clients do. ### Common mistake A public benchmark score tells you little about your task. Test with your own cases and your own standard. ### Related terms - [AI Loop](https://www.excellentaiprompts.com/ai-encyclopedia/ai-loop) - [Prompt Engineering](https://www.excellentaiprompts.com/ai-encyclopedia/prompt-engineering) - [Hallucination](https://www.excellentaiprompts.com/ai-encyclopedia/hallucination) - [Reasoning Model](https://www.excellentaiprompts.com/ai-encyclopedia/reasoning-model) Source: https://www.excellentaiprompts.com/ai-encyclopedia/evaluation (Excellent AI Prompts, updated 2026-11-10) ## What is fine-tuning? Fine-tuning is extra training that adapts an existing AI model to a specific task or style, with a smaller set of examples. ### Key takeaways - Fine-tuning changes the model itself. A prompt changes only one conversation. - It needs hundreds or thousands of good examples and some technical work. - Most business needs are met with good prompts and your own documents, with no fine-tuning. ### How it works A team starts with a trained model and continues the training on its own examples, such as pairs of customer questions and approved answers. The model shifts toward that style and that task. The team then uses the adapted model in place of the general one. ### Example A large insurer fine-tunes a model on 20,000 past claim letters so that new letters match its legal wording. A solo consultant with the same goal gets close by pasting three sample letters into the prompt. ### Why it matters Vendors often offer fine-tuning as a premium service. When you know what it is, you can judge if you need it or if a clear prompt with examples does the job. ### Common mistake Fine-tuning is a poor way to add fresh facts. For facts that change, give the model the documents at the time of the question. ### Related terms - [Training Data](https://www.excellentaiprompts.com/ai-encyclopedia/training-data) - [Retrieval-Augmented Generation (RAG)](https://www.excellentaiprompts.com/ai-encyclopedia/retrieval-augmented-generation) - [Prompt Engineering](https://www.excellentaiprompts.com/ai-encyclopedia/prompt-engineering) Source: https://www.excellentaiprompts.com/ai-encyclopedia/fine-tuning (Excellent AI Prompts, updated 2026-11-10) ## What is generative AI? Generative AI is AI that creates new content, such as text, images, audio, video or code, in answer to a request written in plain language. ### Key takeaways - Generative AI makes new content. Earlier AI mostly sorted or scored existing data. - ChatGPT, Claude, Gemini and Copilot are generative AI tools. - The output is new each time, so two runs of the same request can differ. ### How it works A generative model learns from a very large collection of existing text, images or audio. When you give it a request, it builds an answer one small piece at a time. Each piece is the one the model judges most likely to come next. ### Example A solo business owner asks an AI tool for three versions of a sales email for a new course. The tool writes all three in under a minute, and the owner edits the best one. ### Why it matters Writing, summarizing and drafting are a large part of professional work. Generative AI does the first pass of that work in seconds. ### Common mistake Generative AI can state false things with confidence. Check names, numbers, quotes and sources before you publish or send. ### Related terms - [Large Language Model (LLM)](https://www.excellentaiprompts.com/ai-encyclopedia/large-language-model) - [Prompt](https://www.excellentaiprompts.com/ai-encyclopedia/prompt) - [Hallucination](https://www.excellentaiprompts.com/ai-encyclopedia/hallucination) - [Multimodal AI](https://www.excellentaiprompts.com/ai-encyclopedia/multimodal-ai) Source: https://www.excellentaiprompts.com/ai-encyclopedia/generative-ai (Excellent AI Prompts, updated 2026-11-10) ## What are AI guardrails? Guardrails are the rules and checks that keep an AI system inside safe and approved limits, such as blocked topics, required formats and actions that need approval. ### Key takeaways - Guardrails can sit in the instructions, in the software around the model, or in both. - Typical guardrails limit what the AI may say, what data it may see and what it may do. - Instructions alone are a weak guardrail. Hard limits in software are stronger. ### How it works Some guardrails are instructions in the system prompt, such as "never quote a price". Stronger ones are built into the tool. A filter checks input and output, a permission list limits which tools the agent may use, and an approval step stops risky actions until a person agrees. ### Example A firm puts an AI assistant on its website. The assistant can answer questions from the public help pages. It cannot see customer records, and it must pass any refund request to a human. ### Why it matters Before you let AI talk to clients or act in your accounts, decide what it must never do. Then make those limits real in the tool settings, not only in the prompt. ### Common mistake People rely on a polite instruction to stop a bad outcome. If an error would be costly, remove the ability, or require approval. ### Related terms - [System Prompt](https://www.excellentaiprompts.com/ai-encyclopedia/system-prompt) - [Prompt Injection](https://www.excellentaiprompts.com/ai-encyclopedia/prompt-injection) - [Human in the Loop](https://www.excellentaiprompts.com/ai-encyclopedia/human-in-the-loop) - [AI Agent](https://www.excellentaiprompts.com/ai-encyclopedia/ai-agent) Source: https://www.excellentaiprompts.com/ai-encyclopedia/guardrails (Excellent AI Prompts, updated 2026-11-10) ## What is an AI hallucination? A hallucination is an AI answer that sounds confident and correct but is false or invented, such as a made-up quote, number, source or event. ### Key takeaways - It happens because a model predicts likely text. It does not look facts up. - The risk is highest for exact details such as names, dates, figures, quotes and citations. - You lower the risk when you give the model the source material and ask it to quote from it. ### How it works The model always produces the text that fits best, even when it lacks the facts. If it has no reliable information about a question, it can still write a fluent answer that has the right form and the wrong content. The model gives no sign that this happened. ### Example A lawyer asks an AI tool for court cases that support an argument. The tool returns five cases with full citations. Two of them do not exist. The lawyer finds this only by looking up each case. ### Why it matters One invented fact in a client report or a published article costs trust. A professional who uses AI needs a check step for every factual claim. ### Common mistake Asking the model "are you sure?" is not a check. Use a real source. Give the model the documents, ask for exact quotes with page numbers, or verify the claim yourself. ### Related terms - [Retrieval-Augmented Generation (RAG)](https://www.excellentaiprompts.com/ai-encyclopedia/retrieval-augmented-generation) - [Evaluation (Evals)](https://www.excellentaiprompts.com/ai-encyclopedia/evaluation) - [Human in the Loop](https://www.excellentaiprompts.com/ai-encyclopedia/human-in-the-loop) - [Large Language Model (LLM)](https://www.excellentaiprompts.com/ai-encyclopedia/large-language-model) Source: https://www.excellentaiprompts.com/ai-encyclopedia/hallucination (Excellent AI Prompts, updated 2026-11-10) ## What does human in the loop mean? Human in the loop means that a person reviews, corrects or approves the work of an AI system at set points before the work has an effect. ### Key takeaways - The AI does the volume. The person supplies judgment and takes responsibility. - Put the review where an error would cost the most, such as before sending, publishing or paying. - A review step works only when the reviewer has time and a clear standard. ### How it works You design the process so that the AI stops at a checkpoint and shows its work. A person reads it, fixes it or sends it back, and then approves the next step. Many agent tools have this built in as an approval prompt before an action. ### Example An agent drafts replies to 40 customer emails. It does not send them. The owner scans the drafts in ten minutes, edits three, approves the rest and sends them in one batch. ### Why it matters You stay responsible for what your business sends out. A review step lets you delegate work to AI and still stand behind the result. ### Common mistake A person who approves everything without reading is not a safeguard. Keep batches small enough to review with attention. ### Related terms - [Guardrails](https://www.excellentaiprompts.com/ai-encyclopedia/guardrails) - [AI Agent](https://www.excellentaiprompts.com/ai-encyclopedia/ai-agent) - [Agentic Workflow](https://www.excellentaiprompts.com/ai-encyclopedia/agentic-workflow) - [Hallucination](https://www.excellentaiprompts.com/ai-encyclopedia/hallucination) Source: https://www.excellentaiprompts.com/ai-encyclopedia/human-in-the-loop (Excellent AI Prompts, updated 2026-11-10) ## What is inference in AI? Inference is the moment a trained AI model does its job. It takes your input and produces an output. ### Key takeaways - Training builds the model one time. Inference happens every time someone uses it. - Each answer you get from a chat tool is one inference run. - Inference is what AI providers charge for, usually by the token. ### How it works Your message goes to a data center, where the model runs on specialized chips. The model reads the tokens, computes the most likely next token, adds it to the answer and repeats until the answer is complete. The text you watch appear on screen is that process. ### Example A support team connects an AI model to its help desk. Each customer question that the model answers is one inference call, and the monthly bill adds up those calls. ### Why it matters When you compare AI products or plan a budget, the cost and the speed of inference matter more than the cost of training, which the provider already paid. ### Common mistake A model does not learn from your chats during inference. It answers with what it already has, plus what is in the context window. ### Related terms - [Token](https://www.excellentaiprompts.com/ai-encyclopedia/token) - [Large Language Model (LLM)](https://www.excellentaiprompts.com/ai-encyclopedia/large-language-model) - [Training Data](https://www.excellentaiprompts.com/ai-encyclopedia/training-data) Source: https://www.excellentaiprompts.com/ai-encyclopedia/inference (Excellent AI Prompts, updated 2026-11-10) ## What is a large language model (LLM)? A large language model (LLM) is an AI system trained on a very large amount of text. It reads and writes language by predicting which words come next. ### Key takeaways - An LLM is the engine inside chat tools such as ChatGPT, Claude and Gemini. - It works by prediction. It has no database of facts that it looks up. - Better instructions and better context give better output from the same model. ### How it works During training, the model reads text and tries to predict the next piece of each sentence. It does this billions of times and adjusts itself after each try. The result is a system that can continue any text in a sensible way, which is enough to answer questions, write drafts and follow instructions. ### Example You type a question into Claude. The model reads your full message, then writes the answer one token at a time, and each token is its best prediction for what comes next. ### Why it matters Almost every AI tool a professional uses today has an LLM inside. When you know it predicts text, its strengths and its errors both make sense. ### Common mistake People expect an LLM to behave like a search engine. A search engine finds pages. An LLM writes a likely answer, which can be wrong unless you give it the source material. ### Related terms - [Token](https://www.excellentaiprompts.com/ai-encyclopedia/token) - [Context Window](https://www.excellentaiprompts.com/ai-encyclopedia/context-window) - [Training Data](https://www.excellentaiprompts.com/ai-encyclopedia/training-data) - [Hallucination](https://www.excellentaiprompts.com/ai-encyclopedia/hallucination) - [Multimodal AI](https://www.excellentaiprompts.com/ai-encyclopedia/multimodal-ai) Source: https://www.excellentaiprompts.com/ai-encyclopedia/large-language-model (Excellent AI Prompts, updated 2026-11-10) ## What is machine learning (ML)? Machine learning (ML) is the method behind modern AI. A program learns to do a task from many examples, with no person writing the rules by hand. ### Key takeaways - Machine learning is how AI systems are built. AI is the result you use. - The quality of the examples decides the quality of the system. - You do not need to know the math to use tools built with it. ### How it works A team collects examples, such as millions of emails marked spam or not spam. The program makes a guess for each example, compares the guess with the correct label, and adjusts itself a little. After many rounds, its guesses are good enough to use on new emails. ### Example Your email service uses machine learning to move spam out of your inbox. Your bank uses it to flag a card payment that does not look like your usual spending. ### Why it matters When a vendor says a product uses machine learning, you can ask two useful questions. What examples did it learn from, and how do you measure its errors? ### Common mistake A system learns the errors and the bias in its examples too. Good results on old data do not promise good results on your data. ### Related terms - [Artificial Intelligence (AI)](https://www.excellentaiprompts.com/ai-encyclopedia/artificial-intelligence) - [Training Data](https://www.excellentaiprompts.com/ai-encyclopedia/training-data) - [Large Language Model (LLM)](https://www.excellentaiprompts.com/ai-encyclopedia/large-language-model) Source: https://www.excellentaiprompts.com/ai-encyclopedia/machine-learning (Excellent AI Prompts, updated 2026-11-10) ## What is the Model Context Protocol (MCP)? The Model Context Protocol (MCP) is an open standard that lets AI tools connect to outside apps and data, such as your calendar, your files or your CRM, in one common way. ### Key takeaways - MCP is a common plug between AI tools and other software. - A connection built for one AI tool can work in other tools that support the standard. - Each connection gives the AI real access, so you grant it with care. ### How it works A service such as a calendar or a project tracker offers an MCP server. The server lists the actions it allows, such as read events or create a task. Your AI tool connects to it, sees that list and can call those actions during a conversation, with your permission. ### Example You connect your AI assistant to your calendar and your email through MCP. You ask it to find a free hour next week with a client and draft the invitation. It reads the calendar, picks a slot and writes the draft for your approval. ### Why it matters An AI tool that only chats can advise. An AI tool with connections can act in the apps you already use. MCP is the reason new connections appear so fast. ### Common mistake A connector can read and change real data. Connect only services you trust, give the smallest access that does the job, and keep approval turned on for actions that send or delete. ### Related terms - [AI Agent](https://www.excellentaiprompts.com/ai-encyclopedia/ai-agent) - [Agent Skill](https://www.excellentaiprompts.com/ai-encyclopedia/agent-skill) - [Guardrails](https://www.excellentaiprompts.com/ai-encyclopedia/guardrails) - [Prompt Injection](https://www.excellentaiprompts.com/ai-encyclopedia/prompt-injection) Source: https://www.excellentaiprompts.com/ai-encyclopedia/model-context-protocol (Excellent AI Prompts, updated 2026-11-10) ## What is multimodal AI? Multimodal AI is AI that works with more than one kind of content, such as text, images, audio and video, in the same conversation. ### Key takeaways - A multimodal model can read a photo, a chart, a screenshot or a recording. - You can mix kinds of input, such as a spreadsheet image plus a written question. - Most leading chat tools are now multimodal. ### How it works The model converts each kind of input into the same internal form it uses for text. After that step, it reasons about a picture or a sound clip in the same way it reasons about words, and it can answer in text, and in some tools in images or speech. ### Example A founder photographs a whiteboard after a planning meeting and asks the AI tool to turn it into a task list with owners. The tool reads the handwriting and returns a clean table. ### Why it matters A lot of work lives outside neat text files. Multimodal tools let you start from what you have, such as a screenshot, a PDF scan or a voice memo. ### Common mistake Models still misread small print, complex charts and messy handwriting. Check numbers that a model copied from an image. ### Related terms - [Generative AI](https://www.excellentaiprompts.com/ai-encyclopedia/generative-ai) - [Large Language Model (LLM)](https://www.excellentaiprompts.com/ai-encyclopedia/large-language-model) Source: https://www.excellentaiprompts.com/ai-encyclopedia/multimodal-ai (Excellent AI Prompts, updated 2026-11-10) ## What is a prompt? A prompt is the instruction you give an AI model. It can be a question, a task, a document with a request, or all of these together. ### Key takeaways - The prompt is the main control you have over the output. - A good prompt states the role, the context, the task, the format and how to check the result. - Short prompts get generic answers. Specific prompts get specific answers. ### How it works The model reads your prompt as text and continues from it. Everything in the prompt steers the answer. Details about your audience, your goal and your limits remove guesses the model would otherwise make. ### Example Weak prompt: "Write a LinkedIn post about my course." Stronger prompt: "You write for managers aged 40 and over who want to start a consulting business. Write a 120-word LinkedIn post about my six-week course. Open with a specific problem. End with one question. Do not use hype words." ### Why it matters The same model gives a weak answer or a useful one, and the prompt decides which. Prompt skill is the fastest way to get more value from a tool you already pay for. ### Common mistake People rewrite the same vague prompt again and again. Add the missing context, an example of good output, or a clear format. ### Related terms - [Prompt Engineering](https://www.excellentaiprompts.com/ai-encyclopedia/prompt-engineering) - [System Prompt](https://www.excellentaiprompts.com/ai-encyclopedia/system-prompt) - [Context Window](https://www.excellentaiprompts.com/ai-encyclopedia/context-window) Source: https://www.excellentaiprompts.com/ai-encyclopedia/prompt (Excellent AI Prompts, updated 2026-11-10) ## What is prompt engineering? Prompt engineering is the practice of writing and testing instructions for an AI model so that it gives reliable, useful output. ### Key takeaways - It is a writing skill and a testing skill. It needs no code. - Good prompts are specific about the goal, the audience, the format and the limits. - You improve a prompt by running it, reading the result and fixing one thing at a time. ### How it works You write a first prompt, run it and compare the output with what you wanted. Then you change the prompt. Common changes are more context, an example of good output, a required format, or a step where the model checks its own answer. You keep the version that works and reuse it. ### Example A coach needs a weekly client summary. The first prompt gives long, flat text. The coach adds a sample summary and a rule of five bullet points. The next output matches the sample, and the coach saves the prompt as a template. ### Why it matters A saved, tested prompt turns a task you do each week into a task that takes two minutes. A set of those prompts is a real business asset. ### Common mistake There are no magic words. Clear thinking about what you want does more than any trick phrase. ### Related terms - [Prompt](https://www.excellentaiprompts.com/ai-encyclopedia/prompt) - [System Prompt](https://www.excellentaiprompts.com/ai-encyclopedia/system-prompt) - [Evaluation (Evals)](https://www.excellentaiprompts.com/ai-encyclopedia/evaluation) - [Temperature](https://www.excellentaiprompts.com/ai-encyclopedia/temperature) Source: https://www.excellentaiprompts.com/ai-encyclopedia/prompt-engineering (Excellent AI Prompts, updated 2026-11-10) ## What is prompt injection? Prompt injection is an attack in which text hidden in a web page, an email or a document tricks an AI system into following the attacker's instructions and not yours. ### Key takeaways - A model cannot always tell your instructions from instructions inside content it reads. - The risk grows when an AI agent can browse, read email or take actions. - The defense is limited access and human approval, not a cleverer prompt. ### How it works An AI agent reads outside content as part of its work. An attacker places a sentence in that content, such as an instruction to send the user's contact list to an address. The agent may treat the sentence as a command, because all the text arrives in the same context window. ### Example You ask an AI assistant to summarize your unread email. One message contains hidden white text that tells the assistant to forward your latest invoice to a stranger. An assistant with permission to send email could do it. ### Why it matters As you connect AI to your inbox, your files and your accounts, anything it reads becomes a possible way in. This is the main security question for AI agents today. ### Common mistake No prompt makes a model immune. Give agents the least access they need, keep approval on for sending, paying and deleting, and be careful with content from unknown sources. ### Related terms - [Guardrails](https://www.excellentaiprompts.com/ai-encyclopedia/guardrails) - [AI Agent](https://www.excellentaiprompts.com/ai-encyclopedia/ai-agent) - [Model Context Protocol (MCP)](https://www.excellentaiprompts.com/ai-encyclopedia/model-context-protocol) - [Human in the Loop](https://www.excellentaiprompts.com/ai-encyclopedia/human-in-the-loop) Source: https://www.excellentaiprompts.com/ai-encyclopedia/prompt-injection (Excellent AI Prompts, updated 2026-11-10) ## What is a reasoning model? A reasoning model is an AI model that works through a problem in steps before it gives its final answer. ### Key takeaways - It spends extra time and tokens to plan, check and correct its work. - It suits hard tasks such as analysis, planning, math and complex code. - It is slower and costs more, so it is a poor fit for simple questions. ### How it works A standard model starts to write its answer at once. A reasoning model first writes private working notes. It breaks the task down, tries an approach, checks the result and changes course if needed. Then it writes the answer you see. Many tools show this as a thinking mode you can turn on. ### Example A consultant asks for a pricing model with three tiers and a break-even point for each. A standard model gives quick numbers with an arithmetic error. The reasoning mode takes a minute longer and gets the math right. ### Why it matters Model choice is now a daily decision. Use a fast model for drafts and simple edits. Use a reasoning model when an error would cost you. ### Common mistake More thinking lowers errors but does not remove them. A reasoning model can still argue its way to a wrong answer, so check the result. ### Related terms - [Large Language Model (LLM)](https://www.excellentaiprompts.com/ai-encyclopedia/large-language-model) - [AI Agent](https://www.excellentaiprompts.com/ai-encyclopedia/ai-agent) - [Evaluation (Evals)](https://www.excellentaiprompts.com/ai-encyclopedia/evaluation) Source: https://www.excellentaiprompts.com/ai-encyclopedia/reasoning-model (Excellent AI Prompts, updated 2026-11-10) ## What is retrieval-augmented generation (RAG)? Retrieval-augmented generation (RAG) is a method in which an AI tool first finds relevant passages in your documents and then writes its answer from those passages. ### Key takeaways - RAG lets a model answer from your files, not only from its training data. - It lowers invented answers, because the model writes from real source text. - File upload features, NotebookLM and many company chatbots work this way. ### How it works Your documents are cut into passages and indexed. When you ask a question, the tool searches the index for the passages that match, puts them in the context window with your question, and asks the model to answer from them. Good tools also show which passages they used. ### Example A firm loads its 300-page policy manual into an AI assistant. An employee asks about the travel rule for contractors. The assistant finds the two matching sections and answers with a link to each one. ### Why it matters RAG is how you make a general model useful for your business facts with no retraining. It is the idea behind most products that let you chat with your documents. ### Common mistake RAG is only as good as the search step. If the tool fetches the wrong passage, the model writes a confident answer from the wrong text. Check the cited source. ### Related terms - [Embedding](https://www.excellentaiprompts.com/ai-encyclopedia/embedding) - [Context Window](https://www.excellentaiprompts.com/ai-encyclopedia/context-window) - [Hallucination](https://www.excellentaiprompts.com/ai-encyclopedia/hallucination) - [Fine-Tuning](https://www.excellentaiprompts.com/ai-encyclopedia/fine-tuning) Source: https://www.excellentaiprompts.com/ai-encyclopedia/retrieval-augmented-generation (Excellent AI Prompts, updated 2026-11-10) ## What is a system prompt? A system prompt is a set of standing instructions that an AI tool gives the model before your conversation starts. It sets the role, the rules and the tone. ### Key takeaways - The system prompt applies to every message in the conversation. - Custom GPTs, Claude Projects and Gemini Gems are built around a system prompt that you write. - It is the right place for rules you would otherwise repeat in every chat. ### How it works The tool places the system prompt at the start of the context window, before your first message. The model reads it first and treats it as the standing brief for the session. Your own messages then add the task of the moment. ### Example A consultant creates a Claude Project called Proposal Writer. Its instructions say who the clients are, how the firm prices work and which words to avoid. Each new chat in that project follows those rules with no reminder. ### Why it matters A good system prompt turns a general chat tool into a specialist for your business. You write it one time and reuse it every day. ### Common mistake A system prompt is not a security wall. Users can often get a model to reveal or ignore it, so do not put secrets in it. ### Related terms - [Prompt](https://www.excellentaiprompts.com/ai-encyclopedia/prompt) - [Context Window](https://www.excellentaiprompts.com/ai-encyclopedia/context-window) - [Guardrails](https://www.excellentaiprompts.com/ai-encyclopedia/guardrails) - [Agent Skill](https://www.excellentaiprompts.com/ai-encyclopedia/agent-skill) Source: https://www.excellentaiprompts.com/ai-encyclopedia/system-prompt (Excellent AI Prompts, updated 2026-11-10) ## What is temperature in AI? Temperature is a setting that controls how much variety an AI model puts in its answers. A low value gives steady, predictable output. A high value gives more varied output. ### Key takeaways - Low temperature suits facts, data extraction and code. - Higher temperature suits brainstorming and creative drafts. - Most chat apps set it for you. You control it when you use a model through an API. ### How it works For each next token, the model has a list of candidates with a probability for each. At low temperature it almost always takes the top candidate. At higher temperature it takes less likely candidates more often, so the text varies more from run to run. ### Example A developer builds an invoice reader and sets a low temperature, so the same invoice always gives the same fields. For a tagline generator, the developer sets a higher temperature to get many different ideas. ### Why it matters Temperature explains why the same prompt gives different answers. If you need the same result each time, ask for a fixed format and use a low setting where you can. ### Common mistake A low temperature does not make an answer true. It makes the answer consistent, and a consistent answer can still be wrong. ### Related terms - [Prompt Engineering](https://www.excellentaiprompts.com/ai-encyclopedia/prompt-engineering) - [Inference](https://www.excellentaiprompts.com/ai-encyclopedia/inference) - [Hallucination](https://www.excellentaiprompts.com/ai-encyclopedia/hallucination) Source: https://www.excellentaiprompts.com/ai-encyclopedia/temperature (Excellent AI Prompts, updated 2026-11-10) ## What is a token in AI? A token is the small unit of text that an AI model reads and writes. A token is often a short word or a part of a longer word. ### Key takeaways - Models count tokens, not words or pages. - In English, 100 tokens are roughly 75 words. - Usage limits and prices for AI services are set in tokens. ### How it works Before a model reads your message, the tool cuts the text into tokens. Common words are one token. Rare or long words become two or more tokens. The model then works only with those tokens, both for what it reads and for what it writes. ### Example A 1,500-word report is roughly 2,000 tokens. If you paste it into a chat and ask for a 300-word summary, the request uses about 2,000 tokens of input and about 400 tokens of output. ### Why it matters Tokens explain two things you see in daily use. They set how much text fits in one conversation, and they set what a company pays when it uses a model through an API. ### Common mistake The rule of 75 words per 100 tokens is an estimate for English. Other languages, code and tables often use more tokens for the same content. ### Related terms - [Context Window](https://www.excellentaiprompts.com/ai-encyclopedia/context-window) - [Large Language Model (LLM)](https://www.excellentaiprompts.com/ai-encyclopedia/large-language-model) - [Inference](https://www.excellentaiprompts.com/ai-encyclopedia/inference) Source: https://www.excellentaiprompts.com/ai-encyclopedia/token (Excellent AI Prompts, updated 2026-11-10) ## What is training data? Training data is the collection of examples, such as text, images or audio, that an AI model learns from before anyone uses it. ### Key takeaways - A model knows only what was in its training data, plus what you give it in the conversation. - Training data has a cutoff date. The model does not know later events unless it can search. - Gaps and bias in the data show up in the answers. ### How it works A developer gathers a very large collection of material, cleans it and feeds it to the model during training. The model does not store the documents as files. It keeps what it learned as numbers inside the model. ### Example You ask a chat tool about a law that changed last month. The tool gives the old rule, because its training data ends before the change. When you turn on web search, it finds the new rule. ### Why it matters The training data explains why a model is strong on common topics and weak on your company's private details. For private details, you supply the documents. ### Common mistake People assume the model has read everything and is up to date. Ask the tool for its knowledge cutoff, and supply recent or private facts yourself. ### Related terms - [Machine Learning (ML)](https://www.excellentaiprompts.com/ai-encyclopedia/machine-learning) - [Large Language Model (LLM)](https://www.excellentaiprompts.com/ai-encyclopedia/large-language-model) - [Fine-Tuning](https://www.excellentaiprompts.com/ai-encyclopedia/fine-tuning) Source: https://www.excellentaiprompts.com/ai-encyclopedia/training-data (Excellent AI Prompts, updated 2026-11-10)