Prompt engineering for design is already outdated - make way for context engineering

From prompt engineering to context engineering

Prompt engineering was the star skill of 2023: finding the magic formula, the twist that unlocks the right result. In 2026, this is no longer where quality lies. Models have become good enough that the exact wording of a sentence matters much less than what we now call context engineering - the way in which we structure all the context around the model: system prompt, examples, access to tools, memory management. In agentic workflows, system architecture now accounts for more variance in results than the choice of model itself.

Good news for a designer: this shift brings us closer together, not the other way around. After twenty years of writing creative briefs and specifications, I find in context engineering the same requirements as in my best design documents - defining a context, setting constraints, providing references, guiding towards a precise result. The vague "write some text for a button" prompt produces the generic; the rich context - user persona, brand tone, character constraints, example of the expected result - produces the directly exploitable. Quality remains proportional to investment in framing.

Choosing the right model without getting attached to a version number

Models change every quarter - as I write, OpenAI released GPT-5.5 in April 2026, and Anthropic's Claude range leads the general intelligence index. But as I've explained elsewhere, getting attached to a version number is a mistake: what lasts is tiering logic. Here's how I allocate design tasks on the Claude range, regardless of the decimal of the moment.

On the OpenAI side, the strength of recent GPTs remains their native multimodality: I submit mock-up captures to them for a visual audit. The principle: Opus for thinking, Sonnet for producing, Haiku for volume - and I regularly test the version of the moment on my own cases rather than believing in a ranking.

Permanent context: system prompt, few-shot, caching

The prompt system is the foundation - the equivalent of the permanent brief given to a collaborator at the start of an assignment. It encodes tone of voice, format constraints, preferred vocabulary and forbidden formulations. It remains identical from one call to the next, ensuring a consistency comparable to that of a well-briefed human editor. And with Claude's prompt caching, this permanent context is cached on the server side, reducing the cost and latency of subsequent requests - a detail that makes all the difference when you're dealing with volume.

The most effective technique I know is the few-shot: instead of describing the result in abstract terms, I provide concrete examples.

text System: You're a senior UX writer, B2B SaaS. Your tone is sober, direct and never guilt-inducing.

Examples of error messages we find excellent:

• "This email is already in use. Please log in instead."
• Payment declined. Check expiration date.
• File too large (max 10 MB). Compress and try again.

Task: write 3 error messages in the same register for a password that's too short.

The model captures implicit patterns - length, structure, tone - that no explicit instruction would convey as faithfully.

## From prompt to system: MCP, tool use and structured output

This is where context engineering comes into its own. Rather than copying and pasting responses, I connect the model to my tools. Figma's MCP server, now in general availability, gives the model the direct context of my models - structure, components, variables - without my having to describe everything. And Claude's tool use allows me to define a `generate_component_spec` tool with a precise JSON schema: Claude first reasons about use cases and states, then fills in the schema. The result automatically feeds a specs document into the wiki and a ticket for the dev team. This chain eliminates hours of manual documentation - and it's context engineering, not prompting.

Temperature control completes the picture: low for documentation and specs (predictable results), high for creative exploration (useful variability). I calibrate according to phase - exploration at high temperature, production at low temperature.

## Concrete use cases, from copy UI to mock-up auditing

I use microcopy generation on a daily basis: labels, validation messages, onboarding, notifications. With a calibrated system prompt and two or three reference examples, the output is directly usable in around 80% of cases. Vision-based mockup analysis has transformed my practice: I submit an interface capture and ask for a structured audit - visual hierarchy, typographic consistency, WCAG contrasts, balance of spaces. The model systematically spots problems that the accustomed eye no longer sees, such as insufficient contrast on secondary text. I also use AI to generate a complete palette from a brand color with calculated contrast ratios, and to propose semantic component naming that respects the project's convention.

## Limits, vigilance, evolution

:::callout{type="warning" title="AI is an accelerator, not an oracle"}
Models hallucinate: they produce convincing but false assertions. In design, this results in erroneous accessibility recommendations, invented UX conventions, references to guidelines that don't exist. Every output needs to be validated by human expertise. AI is a gas pedal, not an oracle - especially when it quotes a number or a rule with confidence.
:::

The other limit is disruptive creativity. Models excel at recombining existing patterns and producing quality variations, but the truly original idea, the one that breaks with convention, remains the territory of the human designer. I use AI to eliminate the repetitive and free up time for deep reflection - where my value is strongest. Finally, as the discipline evolves fast, I maintain a versioned library of prompts and contexts, test new models on my real-life use cases, and share findings with the team. Prompt engineering hasn't disappeared - it's become a brick in a larger system, context engineering, and it's this system that a designer must learn to design in 2026.