๐งญ
Prompt Anatomy
The parts of a strong request: purpose, boundaries, requirements, and completion criteria
11 wordsLoading collection...
Loading collection...
Prompt design, context, agents, verification, and safe AI collaboration
LLM fluency is the practical language of working well with modern AI systems. These collections teach how to define an objective, supply context and evidence, specify an output contract, coordinate tools and agents, evaluate results, calibrate permissions, and communicate tone without treating a prompt as a magic phrase. The goal is not jargon for its own sake, but a vocabulary for directing work and checking what comes back.
The parts of a strong request: purpose, boundaries, requirements, and completion criteria
11 wordsPatterns for communicating intent clearly and showing the model what success looks like
9 wordsVocabulary for supplying trustworthy information and managing what the model can use
10 wordsTerms for specifying the structure, detail, readability, and stopping behavior of a response
8 wordsVocabulary for directing multi-step work across people, models, tools, and specialized agents
12 wordsTerms for testing outputs, diagnosing failures, and measuring whether an AI system remains useful
12 wordsVocabulary for treating external content cautiously and limiting what AI systems are allowed to do
8 wordsUseful style words for shaping how an answer reads without changing what it must accomplish
12 wordsComplete vocabulary list for easy reference and copy-paste.
| Term | Definition |
|---|---|
| objective | the specific outcome a prompt asks the model to achieve |
| audience | the intended reader or user whose knowledge, needs, and expectations should shape the response |
| context | the relevant background, conversation, and source material available for interpreting a request |
| scope | the boundary of what a task includes and excludes |
| constraint | a limit or rule the response must respect |
| assumption | a premise treated as true for the task even though it has not been established |
| requirement | a condition the result must satisfy to be acceptable |
| non-goal | an outcome deliberately excluded from a task to prevent unnecessary work |
| deliverable | the concrete artifact or result a prompt asks the model to produce |
| acceptance criterion | an observable condition used to decide whether an output meets a requirement |
| definition of done | a shared checklist of conditions that marks a task as complete |
| Term | Definition |
|---|---|
| instruction hierarchy | the order of authority that determines which instruction wins when directions conflict |
| zero-shot prompting | asking a model to perform a task without giving an example in the prompt |
| few-shot prompting | providing a small number of examples in the prompt to demonstrate the desired behavior |
| in-context example | a demonstration included in the current input that guides a response without changing the model's parameters |
| counterexample | an example that disproves a general claim or shows what the desired output must avoid |
| edge case | an unusual but valid situation near a task's boundaries where normal assumptions may fail |
| delimiter | a character or marker that clearly separates instructions, examples, or data |
| prompt template | a reusable prompt structure with placeholders for task-specific information |
| clarification question | a focused question that resolves missing or ambiguous information before work continues |
| Term | Definition |
|---|---|
| context engineering | designing what information, tools, instructions, and state a model receives at each step |
| context budget | the limited amount of model input available for instructions, conversation, evidence, and outputs |
| source of truth | the designated authoritative record used when copies or claims disagree |
| reference material | documents, data, or examples supplied as supporting information for a task |
| provenance | the origin and history of information, including how it was collected or transformed |
| freshness | the degree to which information is current enough for the task |
| knowledge cutoff | the point after which a model's training knowledge is not expected to include new events or facts |
| grounding | connecting a model's response to supplied, retrieved, or observed evidence |
| citation | an attribution that points readers to the source supporting a claim |
| context compaction | shortening accumulated context while preserving the facts, decisions, and state needed to continue |
| Term | Definition |
|---|---|
| output contract | an explicit agreement about the fields, format, rules, and failure behavior of a response |
| output schema | a formal description of a data structure, including its fields, types, and allowed values |
| structured output | a response produced in a predictable machine-readable shape, often constrained by a schema |
| response verbosity | the amount of detail and explanation in a response |
| reading level | the degree of language complexity expected of a response's reader |
| uncertainty | the degree to which available evidence leaves an answer unknown, ambiguous, or unreliable |
| abstention | a deliberate refusal to answer when evidence, authorization, or capability is insufficient |
| stop condition | a rule that ends generation or an agent loop once a stated state is reached |
| Term | Definition |
|---|---|
| iterative refinement | improving an output through repeated cycles of feedback, revision, and checking |
| task decomposition | breaking a complex goal into smaller parts that can be completed and verified separately |
| checkpoint | a saved or reviewable state where progress is assessed before work continues |
| agent | a model-driven system that pursues a goal through decisions, tool use, and maintained state within set limits |
| tool call | a structured request from a model for the host system to run an external function or service |
| orchestration | coordinating models, tools, data, and workflow steps so they operate as one system |
| delegation | assigning a bounded task and its authority to another person or agent |
| subagent | a secondary agent given a specialized part of a larger task |
| handoff | the transfer of responsibility, context, and work products from one participant to another |
| approval gate | a control that requires explicit permission before a consequential action can proceed |
| human-in-the-loop | a workflow in which a person reviews, guides, or decides part of an automated process |
| Model Context Protocol | an open protocol for connecting AI applications to external data sources, tools, and workflows |
| Term | Definition |
|---|---|
| hallucination | fluent model output that is false, unsupported, or inconsistent with the supplied evidence |
| sycophancy | a model's tendency to agree with a user's stated view instead of giving an accurate independent answer |
| unsupported claim | a statement for which the available evidence does not provide adequate support |
| verification | checking an output against independent evidence, calculations, tests, or authoritative sources |
| eval | a structured test used to measure how well a model or AI system performs a defined task |
| rubric | an explicit set of criteria and scoring rules for judging an output |
| reference set | a curated collection of test inputs and trusted expectations used for evaluation |
| evaluation baseline | a simple or established result used as the comparison point for a new approach |
| model regression | a decline in behavior that previously worked after a model, prompt, tool, or system changes |
| drift | gradual change in inputs, outputs, or conditions that can reduce system performance over time |
| error analysis | the systematic inspection and classification of failures to identify patterns and fixes |
| judge bias | a systematic preference in a human or model evaluator that distorts scores independently of quality |
| Term | Definition |
|---|---|
| prompt injection | an attempt to make a model follow adversarial instructions that conflict with the intended task or policy |
| indirect prompt injection | adversarial instructions hidden in external content that a model reads through browsing, retrieval, files, or tools |
| prompt leakage | the unintended disclosure of hidden instructions, private context, or other prompt contents |
| untrusted content | data from a source that is not authorized to issue instructions and may be inaccurate or malicious |
| guardrail | a policy or technical control that limits unsafe, unauthorized, or out-of-scope behavior |
| security sandbox | an isolated environment that restricts a program's access and limits the impact of mistakes |
| least privilege | the security principle of granting only the permissions needed for a specific task |
| autonomy calibration | matching an AI system's freedom to act with the task's risk, reversibility, and uncertainty |
| Term | Definition |
|---|---|
| concise | brief while retaining the information necessary for the purpose |
| exhaustive | covering all relevant parts or possibilities within a stated scope |
| plain-language | written so the intended audience can understand it quickly and accurately |
| jargon-free | avoiding specialist expressions that the intended audience may not understand |
| technical | using precise domain concepts and terminology for a knowledgeable audience |
| pedagogical | designed to teach through clear sequencing, explanation, and useful examples |
| neutral | even-handed in wording, without advocating a side unless the evidence requires a conclusion |
| empathetic | showing accurate awareness of another person's feelings and perspective |
| diplomatic | tactful about disagreement, risk, or sensitive relationships while remaining clear |
| assertive | direct and confident about needs or conclusions without becoming hostile |
| skimmable | organized so a reader can quickly locate the main ideas and actions |
| actionable | specific enough to guide a concrete decision or next step |