agentleFS
seankoji-comseankoji-com/.agents/skills/choose-design/SKILL.md

choose-design

Skill2 starsChanged 3 days agoMITView on GitHub

Use after product decisions are settled for new UI, requested redesigns or material unresolved visual or interaction choices. For existing UI, preserve its design and limit choices to the requested change; routine edits need no chooser.

The file

We split this skill into the rules it gives an agent

10 rules. About 0 are common, 10 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈10

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    When the existing system determines the edit, implement it directly if requested.

    line 20
  2. Distinctive:

    When a material choice remains, compare only alternatives for that choice.

    line 21
  3. Distinctive:

    Read the request and the code before writing.

    line 42
  4. Distinctive:

    All designers get the same brief, including the request type, existing baseline, allowed changes and what must stay fixed.

    line 59
  5. Distinctive:

    Before dispatch, state the planned count, exact models and verified lineages, with the reason for the count.

    line 81
  6. Distinctive:

    Do not poll, open top-level threads or let a designer start further agents.

    line 96
  7. Distinctive:

    Ask each designer for 2 options, or 3 when the brief supports three distinct directions.

    line 105
  8. Distinctive:

    After the frames, list the labels and names in one line.

    line 151
  9. Distinctive:

    Always wait for the user's choice.

    line 158
  10. Distinctive:

    Stop task-owned delegates when they are done or cancelled.

    line 163

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Choose a visual design

Mockups come from independent designers in different model lineages. Invoking this skill requests that delegation when a material choice remains. The user chooses before anything is built. The coordinator writes the brief, checks the mockups and presents them. It does not design.

Size the run

Classify the request before choosing designers. New UI permits exploration; an explicit redesign permits exploration within its stated boundary. For edits to existing UI, use the current screen as the baseline. Preserve its visual identity, content, components and interactions outside the requested change. A layout, spacing or small feature request alone does not authorise a redesign. When the existing system determines the edit, implement it directly if requested. When a material choice remains, compare only alternatives for that choice. An explicit chooser request keeps this scope. If no choice remains, report that and implement only if requested; do not invent options.

By default, use at least two distinct lineages whenever a material design choice needs this skill. Routine changes determined by the existing system need no chooser. Pick the count from the remaining choice:

  • 2 for a new component, or a change the existing system only partly constrains.
  • 3 for a new screen, page or app, or a redesign with no established system.

A count or explicit model-only roster overrides these defaults. A named model fills one planned slot; fill the remaining slots from the order below. A model mention alone does not increase the count or mean "only this model". For a single unnamed designer, prefer a lineage outside the coordinator's. Label an explicitly constrained single-lineage run honestly. Do not infer that constraint from a small task or familiar visual system.

Write the brief

Read the request and the code before writing. Record the product, the audience, the screen or component, the real content it shows, its states, the target platforms and the existing design tokens, components or screenshots. Ask only about missing intent that would change the design.

A direct invocation, or an authorised workflow selecting this skill, authorises sending the minimal design brief to the selected roster's providers. That brief may contain written intent, design tokens, task-relevant non-secret excerpts and redacted screenshots. User and standing restrictions on providers or content take precedence. Raw files beyond this brief need explicit sharing scope. Record this bounded grant and the invoking request as the authority source in child briefs. Use the sharing-authority procedure in T3 dispatch preflight for T3 delegates. Carry existing consent into child briefs instead of asking again for the same authorised context. Prefer a minimal design brief over raw repository files; host approval and explicit sharing limits still apply.

All designers get the same brief, including the request type, existing baseline, allowed changes and what must stay fixed. For existing UI, include enough visual evidence to preserve it; inspect missing evidence before dispatch or report the gap. Separate requirements from user preferences. Add no style direction beyond these constraints. Leave the unresolved choice to the designers. Do not show one designer another's work.

Pick designers

Apply T3 dispatch preflight before selecting T3 routes. Discover the host's delegation capabilities and exact model identities. Keep any named models and settings. Otherwise go through this order and take the first compatible models, one per lineage: Sol, Opus, Wonder Full, Argon, Grok 4.7. For unnamed slots, skip unavailable or preflight-incompatible routes and continue down the order. A named unavailable or incompatible model remains a conflict under the preflight; do not substitute for it without the user's acceptance. An unnamed route rejected as unavailable or incompatible before the designer starts counts as unavailable; continue down the order. A sharing-authority rejection follows the preflight's sharing procedure, not model substitution to avoid the rejection. A designer that fails after starting is reported, not replaced. Two aliases, versions or effort settings of the same model are one lineage. The coordinator's own lineage is eligible, because the coordinator does not design. Before dispatch, state the planned count, exact models and verified lineages, with the reason for the count. Check permission and sharing compatibility for the whole roster. Start compatible authorised designers without another confirmation. Never fill an unnamed slot with a second model from the same lineage. If the order runs out before the planned count, or a named-model conflict remains, retain partial results and report an incomplete comparison. The user must explicitly accept a smaller set before it counts as the completed run.

Label designers A, B and C in dispatch order. In T3 Code, resolve models with orchestrator_capabilities subject to the host's discovery frequency. Start one delegate_task per designer and record each taskId. Submit independent launches together where supported. Follow the returned next; if it requires yielding before all launches, preserve and resume the pending roster. The first launch or completion does not satisfy the whole roster. Use native delegation elsewhere. Do not poll, open top-level threads or let a designer start further agents.

A delegated coordinator may start designers only when its assignment permits children. Otherwise return the brief and planned roster to the parent. Return all options, HTML, identities and dispatch gaps to the main thread. Only the main thread presents choices and obtains the user's selection.

Designer instructions

Ask each designer for 2 options, or 3 when the brief supports three distinct directions. For existing UI, vary the requested decision while keeping the rest of the baseline consistent across options. Do not broaden the scope to make options look different. For each option, the designer returns:

  • A short, memorable name.
  • A blurb of two to four sentences giving the intent of the design and why it fits the brief.
  • One sentence on how it differs from the designer's other options.
  • One complete, self-contained HTML document of the main screen. It needs inline CSS, no remote fonts, scripts or images, and realistic content in place of lorem ipsum. It must render well at 728px wide. If the real target is wider, the blurb says how the layout changes there.
  • Visible feature IDs and an annotation legend, following report requirements. Read these requirements before briefing designers and apply them when checking and presenting every option.

Designers must not edit the repository. They return HTML in the final report, or write it to a task-owned artifact path the coordinator names and report the path. Designers must not call html_render or rely on a child-thread presentation. A timeout, missing HTML or a bare completion status means no options came back.

Check the options

Open every option and check its controls and annotations against the report requirements. Reject or repair any that breaks the brief, fails to render, loads remote assets or copies a sibling. For existing UI, check each change against the allowed scope. Shared baseline elements are expected, not copied directions. If two designers make the same unresolved choice, ask the later designer for one redo within the same scope. Do not loop beyond that or invent a broader redesign to force variety. Report a designer that failed with its error; do not substitute.

Preserve the design while checking it, except to apply the report's static fallback and note it in the blurb. Correct inaccurate blurbs.

Present

In T3 Code, give each option its own frame, grouped by designer. For each, check the page with html_preview, then show it with html_render at the measured height. Each frame opens with a header in the T3 theme variables. The header holds a label such as B2 · Grok 4.7, the option name and the blurb. Below the header, put the mockup in its own box with its own colours, 16px of padding and var(--radius) corners. Title each frame with its label and name.

After the frames, list the labels and names in one line. Ask the user to pick an option or describe a hybrid using feature IDs, such as "A1-F1 with B2-F3". On hosts without inline HTML, write each mockup to a task-owned file and give the label, model, name, blurb and absolute path for each one.

After the choice

Always wait for the user's choice. Silence or a timeout is not a choice. Record the chosen option or hybrid and the mockup it came from. Build it only if the original request asked for implementation, using the project's real stack and existing components. A further round of options needs a new request.

Stop task-owned delegates when they are done or cancelled. Remove task-owned mockup files after the choice unless the user wants them kept.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.