If you keep asking an AI coding agent to restore a missing action, simplify a screen, or preserve a choice from an earlier prompt, put those decisions in a short project record the agent can access. Define the user and task, state what each view must contain, keep the existing design system, and record consequential choices with their rationale. Then reference that record in later prompts and review the result yourself.
Why the same UI corrections keep returning
A prompt describes what you want now; it may not preserve the context behind earlier choices. Without a shared record, an agent can produce a plausible screen while overlooking who uses it, which actions matter, what information must remain visible, or how the project already handles navigation and styling. Repeated corrections can therefore signal missing shared decisions—not just a need for more emphatic wording.
As an Amazon Associate I earn from qualifying purchases.
A lightweight design decision method can help people and agents work from the same brief. It is a procedure for deciding what a screen needs and why, not a UI library or a guarantee that an agent will remember. The team must store the record somewhere the tool can access and include it in the work context.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →A six-step method for durable UI decisions
1. Establish the context
Before choosing a layout, write down who will use the screen, how often they do the task, the cost of a mistake, and the device or environment they use. Separate confirmed facts from assumptions so an agent does not treat guesses as requirements.
#1 Best Overall
2. Choose the visual system
Keep the project’s existing design system when one is in use. Its components, styles, and patterns give the agent a coherent starting point and help the interface fit the rest of the product. If the project has no system, choose an appropriate platform system; the method’s original article names Material 3 as its web fallback, not as a universal rule.
Using a system does not mean copying a pattern without checking fit. GOV.UK provides examples with code and information about how and when they have been tested, while noting that community-forum ideas may not have been tested. Its guidance encourages teams to conduct their own research when applying examples: GOV.UK Design System and Get started with the GOV.UK Design System.
3. Map objects before layout when it matters
Identify the main objects and their relationships before drawing screens if those relationships affect routes, navigation, or overall structure. If they would not change any of those decisions, skip this step; mapping for its own sake adds work without resolving a design question.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →4. Record choices that are expensive to reverse
Write down decisions that would cause substantial rework if they changed, along with the reason for each. For example, note whether a setting should be visible on the main page or nested in navigation, which actions must stay prominent, and why. This rationale gives a later prompt a basis for preserving the choice or proposing a deliberate change.
Rank #3
5. Specify every view and its states
For each view, state its purpose, required information, and available actions. Include the real loading, empty, and error states as well as the normal state. Otherwise, a design that looks complete in its happy path may omit content or behavior users need when data is missing or a request fails.
6. Deliver and label validation honestly
Give the agent the relevant record with the implementation request, then inspect the output against it. Label validation accurately—proposed, heuristic, automated, manual, or user research—and say what was not tested. Do not describe a screen as user-tested if no users took part, or imply a validation step that did not happen.
Rank #4
Example: stop losing controls on an account-settings screen
Suppose an agent first returns a grid, then changes it after a correction, while important controls disappear in later versions. Before asking for another layout, establish who changes the settings, which changes happen often, which mistakes would be costly, and which objects or account relationships affect navigation. Then record which controls must remain visible, which can be nested, and the reason for those choices.
In a later prompt, provide that record and ask the agent to revise the screen without changing the documented visibility or navigation decisions unless it explains why a change is needed. Check the result against the record and the actual task; a written constraint helps the agent start from shared decisions but does not replace human review.
Best Value
How to compare layout options
When there is more than one plausible design, compare each option against the same questions rather than relying on preference alone:
- Does it fit the user’s task, context, and device?
- Does it contain all required information and actions?
- Are frequent controls easy to find, and are consequential settings handled deliberately?
- Does it remain consistent with the project’s existing design system?
- Does the proposal cover loading, empty, and error states, and is its validation status clear?
These are decision criteria, not a measured ranking of layouts. The DWP Design System recommends starting from standard patterns and components, comparing a proposed variant with the standard in research, and continuing user research even when using standard patterns: How to use the DWP Design System.
Keep the record useful and proportionate
A design record is most useful when it captures decisions the next task actually depends on. It does not need to become a specification for every small edit. For a minor change, a concise prompt may be enough; use the fuller method when missing context or shifting decisions are causing meaningful rework.
Free tools Windows power users keep installed
One-click scans. No signup required.
There is no measured time-saving result for this method: its article states, “Whether it saves time has not been measured.” Treat it as a practical workflow suggestion, not proven productivity advice or a promise that an agent will retain decisions on its own.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.




