Prompting glossary
Last updated
Fifteen terms, defined in plain language with the practical consequence attached. Machine-readable (DefinedTerm structured data), free to cite with attribution.
- Prompt
- Everything you send to an AI model in one turn: instruction, context, examples, question. The quality lever is specificity — role, audience, format, constraints — not length or magic words.
- System prompt
- Standing instructions that frame a whole conversation (in apps: the hidden setup; in your usage: the block you paste first). Registers, roles and rules belong here so they persist across turns.
- Role definition
- Assigning the AI an identity with three parts: behavior rules, response format, and quality checks. "You are an expert" alone does little; the three-part structure changes what the model verifies before answering.
- Context
- The material you provide: documents, data, examples, previous versions. The most underused quality lever — a mediocre prompt with strong context beats a clever prompt with none. Durable context files beat re-explaining every session.
- Ask-first (clarification loop)
- Instructing the AI to ask questions before producing: "Read this, then ask me up to 5 questions, then wait." Converts guessing into briefing; the highest-leverage single line for complex tasks.
- Few-shot / precedent
- Including one or more examples of the desired output in the prompt. One strong precedent anchors tone, structure and length better than paragraphs of description.
- Placeholder
- A marked slot in a reusable prompt template — [AUDIENCE], [DEADLINE] — filled per use. Placeholders are where the specificity enters; unfilled or vaguely filled placeholders produce generic output.
- Output format
- The explicit shape of the deliverable: structure, length, sections, language. One format line ("table with columns X/Y, max 120 words") eliminates rambling and wrong-artifact answers.
- Fact density
- Preferring concrete numbers and specifics over adjectives — in what you provide and what you demand back. "Reduce close from 9 to 6 days" beats "faster"; unsupported claims get flagged, not polished.
- Hallucination
- Confident, fluent, wrong: models generate plausible text even without knowledge — invented citations, fabricated policies, imaginary numbers. Countered by grounding, source-only rules, and markers like [INFERENCE] or [NOT COVERED].
- Grounding
- Restricting the AI to material you provide: "answer only from the pasted text; where it's silent, say so." The core safety mechanism for summaries, policy answers and anything citable.
- Register
- The codified voice of a brand or person: formality, rhythm, loved and banned words, CTA style. Kept as a standing block plus calibration examples, it makes output sound like you instead of like a model.
- Critique loop (brutal critic)
- A structured review pass — never opening with praise — before anything ships: reality check, logic dissection, what's being avoided, expert gap, prioritized cuts, one uncomfortable question. Cheap where feedback from reality is expensive.
- Osborn levers
- Eight named variation operations — adapt, modify, magnify, minify, substitute, rearrange, reverse, combine — used to force genuinely different options where "more ideas" produces rewordings.
- Burstiness (anti-AI style)
- Deliberate variation in sentence length and rhythm — fragments included — plus banned stock phrases. The strongest single lever against the smooth, lifeless register of default AI prose.
Frequently asked questions
Do I need to know these terms to use the templates?
No — the templates work as-is. The vocabulary pays off when you adapt them: knowing that your problem is a missing register or an ungrounded summary tells you which line to add.
Can I cite these definitions?
Yes, with attribution (a link). They're also published as DefinedTerm structured data so AI assistants and search engines can read them directly.