All chapters
  1. Start hereEverything the model reads is text. Only some of it is an instruction
  2. Foundations · 1From a sentence to a prompt: the parts and what each one is for
  3. Foundations · 2Roles and authority: who is allowed to tell the model what to do
  4. Foundations · 3Prompt composition: how an application builds the prompt
  5. Foundations · 4Prompt patterns: five ways to ask, and when each one earns its cost
  6. Surfaces · 5Tool-aware prompting: the tool description is a prompt too
  7. Surfaces · 6MCP prompts and resources: a template you chose versus content that was fetched
  8. Security · 7Prompt injection: when data starts giving orders
  9. Security · 8Defense in depth: assume one layer fails
  10. Operations · 9Long context and prompt placement: where an instruction sits changes what it does
  11. Operations · 10Versioning and provenance: which prompt produced this answer
  12. Operations · 11Testing prompts: golden cases, injection probes and comparisons
  13. Operations · 12Portability: the same prompt on another model is another prompt
  14. Operations · 13Production operations: rolling a prompt out, and rolling it back
  15. Reference · 14Frameworks and tooling: what each one does with your prompt
  16. Reference · 15Advanced topics: other languages, images, hand-offs and attackers who adapt
  17. Reference · 16The insurance lab: one poisoned roof note, end to end
  18. ReferenceGlossary
  19. AboutAbout this site and its sources
Reference

Glossary

Every term this site uses, each with what it is, what it is not, and how it shows up in the insurance example.

Every term this site uses, in one place. Each card opens the same drill-down as the underlined phrases in the text: one line, why it matters here, what it is, what it is not, and how it shows up in the insurance example.

A/B testAn experiment that sends part of live traffic to a new version and compares outcomes with the current one.Adaptive attackAn attack that changes its wording or strategy after each failure, until something gets through.Agent hand-offThe point where one agent's output, often a summary of untrusted text, becomes another agent's input.Application messageText that the application writes at run time, with the builder's authority, such as a case line or a computed result.Assistant messageA message with the role assistant, holding what the model wrote earlier in the conversation.AuthorityThe rank of a piece of text: whether it may direct the model, and over what.Cache prefixThe part of a prompt, from its first token up to the first difference from an earlier request, that prompt caching can reuse.CanaryA small share of real traffic that gets a new version while the rest stays on the old one, so a problem shows up on a few runs first.Chain of thoughtIntermediate reasoning that a model produces before its final answer, either because the prompt asks for it or in a separate thinking phase.ChangelogA short note stored with each version that says what changed and why.Clarifying questionA response that asks for the information needed to answer, rather than guessing it.CompletionThe text a model produces in answer to one prompt.Confused deputyA program that holds some authority and is tricked by a less privileged party into using that authority for the party's purpose.Context windowThe maximum number of tokens a model can handle in one call, counting the prompt and the answer it writes.ContextThe information in the prompt that the model works on, such as a submission, evidence or rules, as opposed to the instructions that say what to do with it.Data markingWrapping or transforming untrusted text so the model can tell where it came from and that it is information, not instruction.DecompositionBreaking a task into steps that are each easier to get right, and handing each step to whatever does it best: a model, code or a person.Defense in depthStacking independent protections so that an attacker must beat all of them, and a failure of one does not become a failure of the whole.Delimiter breakoutAn attack in which data contains the closing delimiter, ends its block early, and continues as if it were outside the data.DelimiterA marker, such as an XML-style tag, placed before and after a piece of data so the model can tell where it begins and ends.Developer messageA message role some providers give to the application developer, ranked above the user and below a platform level.Direct injectionPrompt injection in which the attacker is the person typing to the model.DSPyA Python framework in which you declare a task as typed inputs and outputs (a signature), pick a strategy (a module), and let an optimizer rewrite the instructions against a metric and labeled examples.EscapingRewriting the characters in untrusted text that could open or close a marked block, so the text stays inside it.EvalA run of a system over a set of cases, scored in a way written down before the run, that measures how it behaves.Evidence idA stable identifier, such as EV-ROOF-1101, for one piece of evidence: a roof assessment, a document, a retrieved passage.ExfiltrationAn injection whose goal is to get data out: into a link, a message, a field another party can read.Feature flagA switch read at run time that routes a request to one path or another without a new deploy.Few-shotA prompt pattern that includes a few worked examples so the model can copy the shape and reasoning of a good answer.Golden caseOne input with its expected result, labeled by a person, that every version of a prompt must pass.GuardrailA check in code that runs around the model, on its inputs or outputs or actions, and enforces a rule whatever the model produced.Human approvalA step in which an authorized person reviews and decides, so that no model output has an effect on its own.Indirect injectionPrompt injection through content the application retrieves or receives: a document, a web page, a field, a tool result.Injection probeA golden case built around an injected instruction, checking that the outcome is the same as without it.Instruction hierarchyThe ranking a model is trained to follow when text from different sources disagrees: builder above user, and data below both.InstructionText meant to direct the model, written by someone who has the authority to direct it.InterpolationThe step that inserts a value into a template, producing the text the model will read.IsolationKeeping untrusted text away from the model that plans and acts, by having a separate, tool-less step extract only typed facts from it.JailbreakAn attempt, usually by the user, to get a model to produce something it was trained to refuse.JSON SchemaA standard way to state which fields a JSON value has, their types, which are required and which values are allowed.LangChain prompt templateA class in the langchain_core library that holds message templates with named holes and returns the filled messages when you call invoke with a dictionary.Langfuse prompt managementThe prompt management part of Langfuse, an open-source platform: each prompt has immutable numbered versions, labels such as production point at one of them, and the SDK fetches, caches and fills the prompt.LangSmith promptsThe prompt part of LangChain's LangSmith platform: prompts stored in a workspace, each change a commit with a hash, tags and the Staging and Production environments pointing at commits.Least privilegeGiving each component only the access its job requires, so a compromised component can do little.Lost in the middleThe finding that models tend to use information best at the beginning or end of a long input and worst in the middle.MCPAn open protocol in which a host application connects to servers that offer prompts, resources and tools, using JSON-RPC 2.0 messages.MCP promptA named template a server exposes; the user selects it, for example as a slash command, and the client fetches its messages with prompts/get.MCP resourceA piece of content identified by a URI that the client application decides to read and include as context.MCP serverA program, local or remote, that answers MCP requests from a client inside a host application.MCP toolA function listed by a server with tools/list (name, description, input schema) and run with tools/call when the model asks for it.MessageOne entry in the list an application sends to a model, with a role that says who it is from and content that holds the text.Model parametersSettings such as the model id, temperature and maximum answer length, sent with each request.Model pinningSending a model id that always means the same model, so that a provider release does not change your prompt's behavior overnight.ModelA program trained on large amounts of text that reads a sequence of tokens and produces the text most likely to follow.Multilingual promptA prompt whose parts are in more than one language, or written in one language for input that arrives in another.Multimodal inputAn image, a document page or audio sent to the model with the text, which the model reads directly.Output formatThe part of the prompt that states the exact shape of the answer, such as one JSON object with named fields and allowed values.Output validationChecks, in code, that the model's answer has the right shape and agrees with what the application can compute, before it is stored or acted on.OverrideA decision by the underwriter that differs from the agent's recommendation, recorded with who decided, what, and why.Policy fixtureThe synthetic rule set UW-GUIDE, version 2026.2, effective 2026-07-01, with five rules the eligibility check computes.PortabilityThe ability to run the same prompt on another model or provider and still get the behavior its tests expect.Prompt cachingA provider feature that stores the processed form of a prompt's beginning and reuses it when a later request starts with exactly the same text.Prompt chainingRunning a task as several model calls in sequence, each with its own prompt and the previous call's output as input.Prompt driftA change in what an unchanged prompt produces, caused by something around it: the model behind a name, the serving around the model, the rules, or the inputs.Prompt hashA hash computed over the exact content of a prompt version, which changes if any character of that content changes.Prompt injectionAn attack in which text from someone without authority, placed in the model's input, makes the model act on it as if it were an instruction.Prompt managementThe practice, and the product category, of keeping prompt versions in a store your application reads at run time, with a pointer (a label or tag) that says which version is live.Prompt registryThe place that holds every version of a prompt, refuses edits in place, and names the version production uses now.Prompt templatePrompt text with named holes that code fills with values for each call.Prompt versionOne exact state of a prompt, with its text, variables and parameters, stored once and never edited.PromptEverything the application sends to the model for one call: all messages, tool definitions and attached content.PromptfooAn open-source command-line tool and library in which a test of prompts is a configuration file: prompts, providers, test cases with variables, and assertions.Provenance tagA label, in the prompt or in the record, that names the source of a piece of text and the path it took.Provider adapterA small function per provider that takes the prompt's intent and writes the request body that provider documents.Red teamPeople who attack the system the way a real adversary would, to find what the defenses miss before someone else does.RegressionA failure on a case that passed under the previous version.ReproducibilityThe ability to send exactly the prompt that ran before and compare what comes back.RoleA label on each message, such as system, user, assistant or tool, that says who the text is from.RollbackA return to the version that was active before the current one, done as one action without editing any version.RolloutThe planned move of traffic from the active version to a new one, in stages, with a check and a stop condition between stages.RubricA set of named criteria, each with a definition of what meets it, used to score free-text parts of an answer.Run idA short identifier, such as run-fa74, naming one execution of the underwriting workflow for one submission.Run recordWhat one run of the workflow stored: inputs and evidence, the prompt version and hash, the model, the policy version, the model's output, what code stored, and the person's decision.Self-checkA prompt instruction that asks the model to review its answer against stated criteria before it finishes.Stand-in modelA small program in model_sim.py that reads the composed prompt as text and answers by a printed rule table, so the example runs with no model, key or network.Stop sequenceA short piece of text you send with the request; when the model writes it, generation stops there.Structured outputA request option that gives the provider your JSON schema, so that the answer comes back in that shape.SubmissionOne JSON record with the facts of a quote request, written by an intake step from the broker's email.Success criteriaThe part of the prompt that says what a good answer contains, written so the model can aim at it and code can check it.System promptThe builder's standing instructions, sent in the highest slot the provider offers to customers.TemperatureA number sent with a request that makes the model pick likely words more often (low) or spread its picks more (high).Token budgetThe plan for how many tokens each part of the prompt may use, with room reserved for the answer.TokenA short piece of text, often part of a word, which is the unit a model reads and writes and in which limits and usage are measured.Tool callThe part of a model's answer that names a tool and gives the arguments it chose; a proposal that the application may run or refuse.Tool choiceA setting on the request: let the model decide (auto), require some call, require one named tool, or forbid calls.Tool definitionWhat the application sends to the model about one tool: its name, a plain-text description and a JSON schema for its inputs.Tool descriptionSentences that say what a tool does, when to use it and when not, what each parameter means and what it returns.Tool resultThe data a tool returned, sent back to the model so it can continue.Tool-result launderingInjected text carried in the output of a tool, so that it looks as if it came from the application's own machinery.ToolA function the application offers to the model by name; the model can ask for it, and the application decides whether to run it.TruncationDropping text from a prompt so that it fits the window, either deliberately by policy or by a naive rule such as cutting from the end.Trust boundaryA place in a system where text or data crosses from one author's control to another's, and so must be checked or marked.Underwriter cockpitThe application screen that shows each stored recommendation, its reasons, citations and flags, and records the underwriter's decision.UnderwriterThe person at an insurer who decides whether to accept a risk such as a building, and on what terms.Underwriting agentA program that calls a language model in a loop, runs tools for it, and writes a recommendation for the underwriter.Untrusted dataText written by someone outside the application's control, such as an applicant, a broker, a web page or an outside provider.User messageA message with the role user, holding what the person using the application asked.VariableA named value supplied at run time to fill one hole in a prompt template.Zero-shotA prompt that tells the model what to do without showing any worked example of the answer.

Each term is explained in full on one page and used everywhere else. Where the words come from is on About.