Prompts

A prompt is an intention encoded as input. Precision in = precision out.

The model does not guess your intent — it processes your input. Master the input, own the output.

What Is a Prompt

A prompt is any input that triggers a model response: a question, an instruction, a document, an image, a voice recording, or a combination. The model maps input to output using learned patterns. Your job is to shape the input so the output lands where you intended.

Amplify Your Agency

Domain	Prompt page	Archetype	Outcome	Tools
Analysis	Analysis	The Realist	Insight and pattern	Perplexity, Claude, Gemini
Strategy	Strategy	The General	Direction and plan	Claude, OpenAI o-series
Predictions	Predictions	The Oracle	Calibrated foresight	Superforecaster frameworks
Coding	Coding	The Engineer	Structure and function	Cursor, Copilot, Claude Code
Web Design	Web Design	The Craftsman	Interfaces that work	v0, Bolt, Lovable
Visual Art	Visual Art	The Artist	Composition and style	Midjourney, DALL-E, Flux
Music	Music	The Composer	Feeling with structure	Suno, Udio, Stable Audio
Video	Video	The Director	Motion as meaning	Runway, Sora, Kling
Promo Video	Promo Video	The Advertiser	Conversion in seconds	Runway, Kling, Pika, CapCut
Games	Games	The Designer	Loops that teach	Unity AI, Inworld, Scenario
Voice	Communication	The Translator	Speech and text bridge	Whisper, ElevenLabs, Deepgram
Communication	Communication	The Messenger	Influence and scale	Grammarly, Claude, Jasper
Creative	Creative	The Dreamer	Novelty and vision	All of the above
Library	Library	The Librarian	Copy, paste, iterate	LLMs

Inspiration:

• Github Prompt Libraries

Types of Prompt

Type	Answers	What it sets
Chat prompt	What do I want now?	One-turn task or question
System prompt	Who is processing?	Character, frame, constraints
Agent	Who runs autonomously?	System prompt + model + tools + context
Command prompt	What is being done?	Task, input, expected output
Specification	What must be true at the end?	Acceptance criteria, constraint architecture

An agent is a system prompt promoted to character — see Agents for how they are designed and evolved.

Input Modalities

What you can give an AI model — the intention side:

Modality	Examples	Notes
Text	Instructions, questions, documents	Universal — every model accepts
Code	Functions, repos, error traces	Specialised models apply
Image	Diagrams, UIs, photos, screenshots	Vision models
Audio	Voice recordings, calls	STT first, then reasoning
Video	Clips, screen recordings	Emerging — limited model support
Structured data	JSON, CSV, tables	Inject as text or tool call
System context	Memory, tool state, prior outputs	Context engineering layer

Output Modalities

What you get back — the outcome side:

Modality	Examples	Prompt page
Text / prose	Analysis, copy, strategy, plans	Communication
Code	Functions, tests, scripts, configs	Coding
Images	Visual assets, concepts, art	Visual Art
Audio / speech	Narration, voice agents	Communication
Video	Creative, promo, explainer	Video
Structured data	JSON, tables, reports	Analysis
Actions	Tool calls, API triggers, file writes	Agents

Modality Matrix

Which tool handles each input → output combination:

Input ↓ / Output →	Text	Code	Image	Audio	Actions
Text	Claude	Claude Code	Midjourney	ElevenLabs	Claude + tools
Code	Claude	Cursor	—	—	Claude Code
Image	Claude (vision)	—	Flux / Ideogram	—	—
Audio	Whisper → Claude	—	—	Cartesia	VAPI
Structured data	Claude	Claude Code	—	—	Claude + tools

See the full modality reference for model-level detail.

Techniques

Technique	What it does	Best for
Zero-shot	Direct ask, no examples	Clear tasks with known output format
Few-shot	Provide 2–3 input/output examples	Pattern replication
Chain-of-thought	"Think step by step" prefix	Multi-step reasoning
Role / persona	Declare who is processing	Consistent frame across tasks
System prompt	Define character + constraints	Autonomous or repeated tasks
Negative prompting	State what NOT to produce	Image generation, creative control
Specification	Full intent contract with acceptance criteria	Agent delegation, complex tasks
Iterative refinement	Build through feedback loops	Drafting and editing
Context stuffing	Load relevant documents as context	Long-horizon or domain-specific tasks
Decomposition	Break complex tasks into sub-tasks	Multi-step production pipelines

Prompt Disciplines

As models move from chat to autonomous agents, prompting fractures into four disciplines:

Discipline	What it is	Goal
Prompt Craft	Clear instructions, guardrails, examples	Reliable single-turn responses
Context Engineering	Curating the optimal token set (tools, docs, memory)	Comprehensive information for autonomous tasks
Intent Engineering	Encoding purpose and decision boundaries	Aligning agents with strategy
Specification Engineering	Agent-fungible documents agents execute against	Embedded oversight without human intervention

Specification Primitives

For agent-grade specifications, five elements are required:

1. Self-contained problem statement
2. Acceptance criteria
3. Constraint architecture
4. Decomposition
5. Evaluation design

Tool Selection

Which tool for which job — first choice is default, second is fallback:

Modality	JTBD	1st Choice	2nd Choice	Prompt page
Text — Reasoning	Think through hard problems	Claude	OpenAI o-series	Analysis
Text — Research	Find and synthesise answers	Perplexity	Gemini Deep Research	Strategy
Text — Writing	Draft, edit, persuade	Claude	Writer.com	Communication
Code	Build and ship software	Claude Code	Cursor	Coding
Web/UI	Design interfaces from prompts	v0	Bolt / Lovable	Web Design
Image	Generate visual assets	Midjourney	Flux / Ideogram	Visual Art
Video — Creative	Motion as meaning	Runway	Kling / Sora	Video
Video — Promo	Conversion in seconds	Arcads	Reel Farm / CapCut	Promo Video
Music	Feeling with structure	Suno	Udio	Music
Voice — TTS	Text to natural speech	ElevenLabs	Cartesia	Communication
Voice — STT	Transcribe speech to text	Whisper	Deepgram	Communication
Voice — Agents	Conversational AI on phone	VAPI	Bland AI	Communication
3D	Generate 3D assets from text/image	Tripo	Meshy	—
Games	Interactive loops that teach	Unity AI	Inworld	Games

First Principles

Five rules that apply across every modality — text, voice, image, music, video, code:

Principle	What it does	Example
Context	Ground the model in your world	"You are a senior infrastructure engineer reviewing a Terraform plan"
Constraint	Narrow the output space	"Under 200 words, bullet points only, no speculation"
Example	Show, don't describe	"Input: X → Output: Y. Now do Z."
Iteration	Refine through feedback loops	"That's close. Now make it more concise and remove the passive voice"
Structure	Shape the response format	"Use XML tags: <analysis>, <recommendation>, <risk>"

Chain-of-thought = Structure + Iteration. Persona prompting = Context + Constraint. Negative prompting = Constraint applied to pixels.

Context

• AI Modalities — Full input/output model matrix with model-level detail
• Agents — System prompts promoted to character: design, roles, autonomy spectrum
• Prompting Capability — The framework across all modalities
• Persuasion — Rhetoric: ethos, logos, pathos, kairos, topos
• Deterministic vs Probabilistic — How prompts compress probabilistic space into deterministic action

Links

• Anthropic Prompt Library
• OpenAI Prompt Engineering
• Learn Prompting
• Prompt Engineering Roadmap
• Gemini Prompt Gallery
• Anthropic XML Guide

Questions

• If input precision determines output quality, which input modality are you least precise with — and what does that cost per run?
• Zero-shot gets speed; few-shot gets accuracy; specification gets delegation. Which step are you stuck at?
• The modality matrix shows what each tool handles — which cell in your workflow still has no tool assigned?
• When does a system prompt become an agent — and where is that line in your current setup?