You use one assistant to organize usability notes for a settings flow. Another is better at challenging the interaction and drafting concise interface copy. The work sits in a visual board with twelve frames, but the first chat refers to “the blue version” and “the screen after save.” Those phrases will not survive a handoff.
The next assistant needs an indexed design brief and access to the artifact, not your first model’s memory of it.
Split research, critique, and copy
Give each assistant a bounded design question:
- Research synthesis: organize permitted observations, quotes, and patterns while preserving their source and limits.
- Critique: inspect a named flow against user goals, system constraints, accessibility, and failure states.
- Interface copy: draft labels, help text, errors, and confirmations for the approved interaction and voice.
Keep evidence separate from interpretation. “Three participants missed the disclosure” is an observation only if you checked the record. “The hierarchy is weak” is a design judgment. A fluent critique should not erase that distinction.
Make visual references precise
RelayLink carries text. It does not move design files, screenshots, prototypes, whiteboards, cursor position, or browser state. A useful briefing points to the source artifact with:
- file, page, flow, and frame names or identifiers;
- an artifact URL and version or review timestamp;
- the user goal and scenario;
- research findings with approved source references;
- platform, viewport, content, and technical constraints;
- decisions and rejected directions;
- known accessibility questions;
- the exact frames and states to review;
- the output wanted next.
Describe essential facts that the artifact alone does not show, such as what happens after a failed save or which content is user-generated. Do not attempt to recreate every pixel in prose. The destination assistant needs its own authorized access to the board or prototype. A URL in the package grants no permission.
Hand the design state to yourself
Connect both assistants to the same RelayLink account. Address the package to your own email, then draft, review, and confirm it. The model never sends autonomously. Check frame references, research attribution, private participant details, product secrets, and any claim that users preferred a direction.
The self-addressed package creates no contact request, email, magic-link page, or second account. It remains unread until an assistant on the account pulls it, and the account portal is the fallback. RelayLink does not synchronize either model’s context. It carries only your selected, approved briefing. The self-handoff guide explains the loop.
Ask for critique you can inspect
“Review the design” often produces a generic checklist. Name the concern: trace keyboard focus through the confirmation flow, compare error recovery in frames A12 through A16, or test whether the copy distinguishes save from publish.
Treat the response as another opinion. The model cannot establish usability by describing a plausible user. Verify important questions with the artifact, accessibility tools, technical owners, research, and the people affected. If two assistants disagree, preserve the reasons and decide what evidence resolves them.
When this workflow is a poor fit
Keep the task with the assistant that can see the artifact when the work depends on spatial relationships, animation, precise visual contrast, or rapidly changing comments the other assistant cannot access. A text handoff cannot substitute for the visual source.
Do not move participant identities, unreleased designs, client work, or protected research to an unapproved provider. Employer, client, consent, and provider rules still apply. If the safe text and permitted artifact access are not enough for a meaningful critique, stop rather than asking the destination model to invent what it cannot see.