What is Prompt Engineering in AI Notes?

Prompt engineering is the systematic discipline of designing, structuring, and optimizing textual instructions provided to large language models to maximize output accuracy, consistency, and structural compliance.

In casual conversation, prompt engineering involves asking clever questions. But in production software development, prompt engineering is rigorous software engineering: designing robust templates, enforcing schema validation, preventing adversarial injections, and eliminating non-deterministic hallucinations.

Core architectural principles of effective prompt design

High-reliability meeting summarization prompts rely on four foundational techniques:

1. Role and persona specification

Clear framing constrains the model’s semantic latent space:

You are a senior executive technical scribe. Your sole objective is to
extract verified engineering decisions, action items, and structural risks
from meeting transcripts.

2. Strict negative constraints

Generative models tend to over-explain or add conversational fluff (“Sure, here is your summary:”). Negative constraints suppress unwanted tokens:

• “Never output conversational preamble or pleasantries.”
• “If a deadline was not explicitly confirmed with a date, mark it as ‘Tentative / Unconfirmed’.”
• “Do not invent or extrapolate; rely exclusively on verbatim transcript evidence.”

3. Schema and format enforcement

Instructing the model to emit pure GitHub Flavored Markdown ensures seamless integration with note-taking editors:

Output the summary strictly in the following Markdown format:

## Key Decisions
- [Decision 1] (Timestamp)

## Action Items
- [ ] [Task] — **Owner:** @name — **Due:** YYYY-MM-DD

4. Few-shot in-context demonstrations

Providing two realistic, high-quality examples of raw transcripts and corresponding ideal outputs dramatically reduces format deviations across open-weights local models (like Mistral and Llama).

Prompt architecture in Oyma

In Oyma on macOS:

• Summary prompts are tailored to specific meeting archetypes: Daily Standups, Incident Post-Mortems, Customer Discovery, Architecture Reviews, and 1-on-1s.
• Prompts instruct local models to preserve technical terms, code snippets, and exact participant names.
• All generated output is clean CommonMark that seamlessly renders inside Oyma‘s live-preview editor and your existing Obsidian vault.

Controlling sampling parameters: Temperature and Top-P

Prompt engineering extends beyond prompt text into model inference hyperparameters:

• Temperature (0.0 – 0.2): Keeps sampling near-greedy, ensuring that decisions and action items are extracted deterministically rather than creatively.
• Top-P (Nucleus Sampling): Restricts the candidate token pool to the most probable cumulative mass, preventing ungrounded linguistic drift.
• Frequency and Presence Penalties: Balanced to prevent repetitive phrasing while ensuring key technical terms are retained across summary paragraphs.

By combining calibrated hyperparameters with rigorous prompt framing, local language models produce executive records of exceptional clarity and truthfulness.