Context
Give Context
Context is the difference between a generic answer and an answer that fits your real situation.
Add context to improve performance
Anthropic Claude Platform Docs
Key Facts
- Context means the goal, audience, source material, constraints, examples, and non-goals around the task.
- Paste or summarize the material the answer should use when accuracy depends on specific facts.
- Examples are often clearer than adjectives when you need a particular tone, style, or format.
Context is not extra decoration. It is the working material. A model can produce a plausible answer from a thin prompt, but it cannot know your audience, standards, source documents, deadlines, preferences, or boundaries unless those details are in the conversation.
The easiest way to decide what context to provide is to ask: "What would a smart person need to know before doing this well?" If the answer is in your head, your notes, your inbox, a document, or a policy, put the relevant part in the prompt.
The context worth adding
- Goal context: what success looks like. "I want a draft I can send today" is different from "I want rough ideas."
- Audience context: who will read or use the output. A CFO, a fifth grader, a hiring manager, and a friend need different answers.
- Role context: the lens you want applied. Editor, tutor, analyst, recruiter, skeptical reviewer, or project manager.
- Source context: the material the answer must use. Paste the relevant text when accuracy matters.
- Constraint context: length, tone, format, deadline, reading level, compliance needs, and things to avoid.
- Example context: a sample of what good looks like. Examples are especially helpful for style, structure, and edge cases.
- Non-goal context: what you do not want. "Do not rewrite the quotes" or "do not suggest paid tools" prevents common drift.
Use source material when the answer depends on facts
If you ask, "Summarize the policy," the model needs the policy. If you ask, "What did the meeting decide?" it needs the notes or transcript. If you ask, "Does this contract clause look risky?" it needs the clause and the standard you care about. Without source material, the model can only answer from general patterns.
Source-grounded prompt: Use only the text below unless you clearly label a point as an inference. Summarize the decision, open questions, and next actions for a manager who missed the meeting.
Examples beat adjectives
"Make it professional" can mean crisp, formal, warm, executive, legalistic, or bland. If style matters, include a small example. You can say: "Match the direct, friendly style of this paragraph" and paste the paragraph. Anthropic's guidance treats examples as one of the most reliable ways to steer format and tone, and Google lists few-shot examples as an optional prompt component for more controlled outputs.
Output shape is context too
Models often default to a general essay-like answer. If you need a table, checklist, JSON object, executive summary, lesson plan, rubric, or rewrite-only response, say so. OpenAI's length guidance recommends asking for the exact length or shape you want. This is not micromanaging. It is part of the job description.
How much context is too much?
Too little context causes guessing. Too much unfocused context causes dilution. Give the material needed for the task, then tell the model what to prioritize. If you paste a long document, ask for a first pass that extracts the relevant sections before asking for the final answer.
Useful line: "First identify which parts of the context matter for this task, then answer using only those parts."