Prompting
What happens when the quality of your questions determines the quality of your workforce?
Questioning is a human capability. Prompting is questioning applied to intelligence systems — biological, phygital, or both. The same principles apply: clarity of intention, structured context, feedback that compounds.
Capability, Not Skill
| Frame | Implication |
|---|---|
| Prompting as skill | Learn the syntax, use the tool, move on |
| Prompting as capability | Shape how intelligence flows through every system you touch |
Skills get automated. Capabilities direct automation. The person who can prompt well isn't "using a tool" — they're conducting an orchestra of intelligence across modalities.
The Interface Stack
Prompting is the interface layer between intention and execution. It operates across every modality:
| Modality | What You Prompt | What Returns |
|---|---|---|
| Text | Questions, context, constraints | Reasoning, analysis, code |
| Image | Layout, palette, style, negatives | Visual composition |
| Video | Beat sheet, sequence, motion | Moving narrative |
| Voice | Tone, persona, pacing | Spoken presence |
| Code | Intent, architecture, tests | Working software |
| Agents | Goals, tools, guardrails | Autonomous action |
Same capability. Different outputs. The skill is modality-specific syntax. The capability is structured communication of intent.
Prompt Architecture
Every effective prompt — text, visual, agent — has the same skeleton:
CONTEXT → What does the system need to know?
INTENT → What outcome do you want?
CONSTRAINT → What must NOT happen?
FORMAT → What shape should the output take?
FEEDBACK → How will you measure and iterate?
| Missing | What Fails |
|---|---|
| Context | System guesses your situation wrong |
| Intent | Output is technically correct but useless |
| Constraint | Unwanted elements, wrong tone, scope creep |
| Format | Right content in wrong shape |
| Feedback | First attempt becomes final attempt |
Weak vs Strong
| Weak | Strong | What Changed |
|---|---|---|
| "Write me a marketing email" | "Write a 150-word email to SaaS CTOs who just hit 50 employees. Tone: direct, no fluff. CTA: book a demo. Avoid: buzzwords, 'leverage', 'synergy'." | Context, constraint, format |
| "Make a nice diagram" | "Information design poster, 16:9, dark background #111827. Three-column layout. Left: five stacked bars. Centre: ten nodes in a clockwise ring. Right: single block with label." | Spatial specificity |
| "Build me an agent that does research" | "Agent with web search + file read tools. Goal: find the 3 most-cited papers on X from 2024. Output: markdown table with title, authors, citation count, one-sentence finding. Stop after 3." | Tools, goal, format, guardrail |
The strong versions aren't longer for the sake of length. They're specific where vagueness would cost you iterations.
The Feedback Loop
Prompting is not one-shot. It's a control system.
PROMPT → OUTPUT → EVALUATE → ADJUST → PROMPT
↑ |
└────────── each cycle sharpens ───────┘
| Cycle | What Improves |
|---|---|
| 1 | You discover what the model misunderstood |
| 2 | You add the constraint that was missing |
| 3 | You refine format to match actual need |
| 4+ | You have a reusable prompt template |
Most people stop at cycle 1. The capability compounds at cycle 3+.
By Modality
Each modality has specific technique. The capability transfers; the syntax doesn't.
| Modality | Technique Page | Key Lesson |
|---|---|---|
| Text | Prompting | Structure beats length. XML tags, role anchoring, chain-of-thought. |
| Visual | Visual Prompting | Describe WHERE things go, not just WHAT they are. Negative prompts. |
| Code | AI Coding | Intent + architecture + test expectations. |
| Agents | Agent Frameworks | Goals + tools + guardrails + stop conditions. |
Practice Protocol
- Prompt, then diagnose — When output is wrong, find which skeleton element was missing
- Save your best prompts — Build a library. Templates emerge from repeated wins.
- Vary one thing — Change context OR constraint OR format. Never all at once.
- Read the model's misunderstanding — The error tells you what your prompt actually said vs what you meant
- Cross-modality practice — Prompt text, then image, then video. The capability transfers.
The Shadow
Over-prompting. Writing 2000-word system prompts when 50 words would do. Prompt engineering as procrastination — perfecting the instruction instead of shipping the output. Treating models as fragile when they're robust to reasonable variation.
Compound Effect
| Prompting + | Creates |
|---|---|
| Questioning | Better diagnostic prompts — finding root causes faster |
| Writing | Clearer instructions — fewer iterations to usable output |
| Visualisation | Spatial prompts that produce what you pictured |
| Systems Thinking | Multi-step agent prompts that handle complexity |
| Taste | Knowing when the output is good enough vs needs another pass |
Context
- Prompts — Text prompting technique, provider guides, copy-paste templates
- Visual Prompting — Image and video prompt structure
- Mantra — When prompts become internal triggers
- Questioning — The human skill underneath
- AI Coding — Prompting applied to software
- LLM Models — Which models respond to what