The working techniques, with real examples

The XML prompt structure

Plain paragraphs of instructions get harder for a model to parse correctly as a prompt grows — it has to guess where one instruction ends and another begins. Wrapping each logical section in a named tag removes that ambiguity, because the model can see exactly where your role definition ends and your task begins.

The pattern used across the frontier labs' own documentation: role → context → task → instructions → constraints → output format, each in its own tag.

Why this works: models are trained heavily on structured/marked-up text, so they parse tagged sections with much higher fidelity than an equivalent wall of prose — fewer missed instructions, less drift on long prompts.

The "magic prompt" (meta-prompting)

Instead of writing the perfect prompt yourself, you ask the model to write (or improve) the prompt for you. You describe your goal in plain, messy language, and the model returns a polished, structured prompt you can reuse.

The model analyzes your examples and hands back a structured prompt — you're using the model to build your own tools instead of trial-and-erroring the phrasing yourself.

The 5-layer context framework

Beyond a single prompt, a durable AI setup needs layered context, each answering a different question:

• Identity / role — who is the model acting as? ("a senior tax attorney," not "a lawyer")
• World / background context — what does it need to know about your situation that it can't infer?
• Task — what specific outcome are you asking for right now?
• Instructions — the step-by-step method for getting there
• Constraints — the hard boundaries: what NOT to do, format limits, tone limits

Most one-off prompts only specify task and skip the rest — fine for a single question, but it breaks down the moment you want a reusable assistant that behaves consistently across many conversations.

Image prompting: comma-separated keywords

Image prompting is structurally different from chat prompting: instead of full sentences, you list elements separated by commas, like a creative director calling out what's in a shot. The standard element order: subject → action → environment → atmosphere → camera → lighting, and whatever you place first gets the most visual weight.

The one critical keyword: without an explicit style instruction like "ultra-realistic, cinematic style," the model may default to a cartoon, a 3D render, or a random aesthetic — always specify the visual register you want.

Camera angle vocabulary worth memorizing: a Dutch angle (tilted, for tension), a low angle (looking up, for power), a cowboy shot (torso-up framing), a 100mm lens (extreme close-up) versus a 15–20mm lens (wide, good for interiors/landscapes).

Video prompting: natural language, not keywords

Video models need to understand motion, timing, and narrative — so you switch back to full sentences describing what happens, in order.

Image-to-video beats text-to-video for control: uploading a reference image as the first frame gives the model a concrete visual anchor instead of a random starting point. Providing both a start frame and an end frame is even stronger — the model generates the transition between two states you specified.

Character consistency discipline: reuse the same reference headshot across every prompt in a scene, and describe wardrobe explicitly and identically every time ("white t-shirt," never just "shirt").

The professional workflow, honestly: validate a prompt on a cheap model at low resolution first, then re-run the winning prompt on a premium model at full resolution for the final. Expect to iterate a prompt many times before a shot works.

Four rules for designing an agent that actually works

• Scope ruthlessly. "Personal assistant" fails as a project — too broad. "Weekly expense-report categorizer" ships, because the narrower the scope, the higher the reliability.
• Give tools, not instructions. Don't script every step — expose the right tools and let the model figure out the sequence.
• Plan for failure explicitly. Every production agent needs a human-in-the-loop checkpoint before anything irreversible, a max-cost ceiling, a max-step limit, and a rollback path.
• Judge by operational readiness, not benchmark scores. Measure time-to-resolution, error rate, and real user satisfaction, not leaderboard rank.

The "AI employee" pattern (multi-tool personal agent)

A single-purpose chatbot answers questions. A multi-tool agent is wired into several tools at once — email, calendar, a voice layer for phone calls, and web search — and given a standing goal rather than a single instruction, so it can decide which tool to use at each step.

Build pattern in practice: connect the agent platform (n8n, Gumloop, Lindy) to each tool via its native integration or an MCP connector, give the agent a clear escalation rule ("if a request involves money, pause and ask me first"), and test it on read-only tasks before letting it take an action that can't be undone.

The core loop, and the one rule that prevents disaster

The loop: describe what you want in plain English → the tool generates code → you preview → you describe what to change → repeat. What separates a working app from an unmaintainable mess is one discipline: write a one-page PRD before your first prompt.

A minimal PRD answers three questions before any code exists: what problem are you solving, who is it for, and what does "done" look like. Thirty minutes here routinely saves hours of "fix one bug, introduce two more" later.

Non-negotiables regardless of tool: commit every working state to version control before your next experiment; never let API keys get hardcoded into committed files; don't ship anything you can't explain in one sentence per function.