daBongo LMS AI Training Courses

Lesson Content

The context window – Copilot's working memory.

When you interact with Copilot, everything in the current conversation – your messages, Copilot's responses, and any documents you have shared – exists in what is called the context window. The context window is Copilot's working memory for the current session: it processes everything in it to generate each new response.

Practical implications:

• Earlier messages in a long conversation can affect later responses (Copilot can refer back to context you established earlier)
• When a conversation becomes very long, earlier parts of it may be treated as less prominent in generating new responses
• Starting a new conversation resets the context entirely

Calibration signals – how Copilot knows what you need.

Copilot uses every element of your prompt as a calibration signal – information that tells it what kind of response to produce. Common calibration signals include:

• Role signals: "I am a financial analyst…" -> calibrates vocabulary, depth, and framing
• Audience signals: "Write this for a non-technical leadership team…" -> calibrates language complexity
• Tone signals: "Keep this professional but warm…" -> calibrates register
• Format signals: "Format as a bulleted list…" -> calibrates output structure
• Constraint signals: "Under 150 words, avoid jargon…" -> calibrates length and vocabulary

Every element of the four-part opening message framework (from Part 1) functions as a calibration signal. This is why richer prompts produce better responses – not because more words are better, but because more calibration signals reduce ambiguity.

Why Copilot produces different responses to similar prompts.

Language models have a degree of natural variability in their outputs – the same input does not always produce the exact same response. Additionally, small differences in wording – words that seem synonymous to humans – can produce meaningfully different responses because they activate different patterns in the model's training.

Practical implication: if a prompt produced a great response, save it. Do not assume the same prompt will produce the same response next time – it usually produces something similar, but not identical.

What types of prompts produce the most consistent results.

• Highly specific prompts with explicit calibration signals
• Prompts that provide the content to work with (draft to improve, text to summarize) rather than asking Copilot to generate from scratch
• Prompts with specific format requirements that constrain the output structure
• Prompts that ask Copilot to evaluate or analyze rather than create from nothing