Product Design
Decide what the interface should do, then route who builds and verifies it: pick the right interaction, make scope and consequence clear, cover reality beyond the happy path. This skill owns the decision; it routes the build, verification, and copy out (ownership map in Related skills).
- IS: the decision layer. From a brief, spec, mockup, intent, or existing UI: choose the right interaction and control, name the object, scope, and consequence of actions, enumerate every reachable state, set resilience expectations, require accessibility as task completion. It decides, then routes build and verification out.
- IS NOT:
- building or styling UI, visual direction, palettes, type: use
ui-design.
- auditing the built result (rendered quality, a11y markup, keyboard, layout, performance, type surface, React/Next code-level UX with a ship verdict): use
ui-audit.
- copy wording, persuasion, or AI-ism removal: use
copywriting.
- deep typography or motion: use
typography-audit or ui-animation.
product-design or ui-audit?
Dispatch on the artifact, not the topic. Both care about states and interaction; they act at different moments.
| You have... |
The question is |
Use |
| A brief, spec, mockup, intent, or a UI you decide about |
What should exist: the right interaction, the action's name, which states should be reachable |
product-design |
| Code, a diff, or a running UI you decide on |
Is the built result right: states covered, accessible, renders and behaves correctly, ready to ship |
ui-audit |
One artifact often needs both in sequence: product-design decides the states that must exist, ui-audit verifies the built code and rendered result implement them. Looking at code or a running screen is ui-audit's turn. This skill reviews the decision and stops at decision altitude; it never writes line-level code fixes.
Operating contract
- Cite a stable rule ID for every finding or non-mechanical decision. Never invent an ID; if none fits, record a coverage gap.
- The project's design system and
AGENTS.md outrank this skill's defaults. Defer to them.
- Never restyle or rebuild. Decide, then route the build to
ui-design.
- One mode per request, resolved from the user's verb before acting.
Request modes
Resolve the mode from the user's verb and artifact, then load only that mode's references.
| Mode |
Dispatch when the user asks for |
Load |
| shape (default) |
"design the flow for", "what control here", "how should this work", "is this the right pattern", a brief with no settled UI |
references/product-judgment.md, references/surfaces.md |
| spec |
"spec the right interaction", "define the expected states", judgment applied before or during a build |
references/surfaces.md, references/naming-and-copy.md; route the build to ui-design |
| review |
"review this for product correctness", "what's wrong with this UX decision", "audit this flow" |
references/interface-quality.md, references/rules.md |
| action |
"what should this action affect", "which object or scope does this action cover", or action reversibility is unsettled |
references/naming-and-copy.md; route final wording polish to copywriting |
| harden |
"make this resilient", "what breaks here", error, permission, offline, and destructive paths |
references/surfaces.md, references/interface-quality.md |
Modes chain: shape leads into spec; review leads into harden. When intent is ambiguous, use the narrowest mode the verb supports. A URL, screenshot, route, or component identifies scope; it does not authorize edits.
A material decision: see references/product-judgment.md.
Decision authority
Conflict order, highest first:
- The user's explicit goal and constraints.
- Verified user and product evidence, and what the system actually does.
- Project-canonical guidance:
AGENTS.md or CLAUDE.md, the project's design system, routed sibling skills.
- Sibling-skill ownership: route, do not duplicate (ownership map in Related skills).
- This skill's product design standards (below).
- General interface and platform conventions.
When a request spans authorities, name the owning skill and hand off.
Workflow
Product design pass:
- [ ] Step 1: Classify the request into one mode
- [ ] Step 2: Locate authority (user constraints, project design system, AGENTS.md)
- [ ] Step 3: Load only that mode's reference files
- [ ] Step 4: Identify object, scope, and consequence of each action in scope
- [ ] Step 5: Enumerate reachable states; check coverage (surfaces.md)
- [ ] Step 6: Apply standards; cite a stable rule ID per finding or decision
- [ ] Step 7: Emit output (review and harden use P0-P3); route follow-on work to siblings
For shape, spec, harden, or any material product or flow change, write the compact internal brief in references/product-judgment.md before proposing UI: user, job, current behavior, desired outcome, success signal, non-goals, object, action, consequence, reversibility, permissions, and open decisions.
Product design standards
Five pillars. Each cites its governing rule IDs in references/rules.md and its reference.
- Right interaction. Pick the control from the choice's shape; keep options visible and reversible; prefer inline disclosure over a modal; choose the smallest coherent intervention.
rule/control-matches-cardinality, rule/navigation-vs-action, rule/inline-before-modal, rule/smallest-intervention. See references/product-judgment.md.
- Action naming. Name the object, scope, and consequence; destructive CTAs use Verb plus Noun, never "Confirm" or "OK"; make friction proportional to impact and offer undo when honest.
rule/name-object-scope-consequence, rule/destructive-names-action, rule/destructive-proportional. See references/naming-and-copy.md.
- State coverage. Design every reachable state, not just the populated one; empty states name the object and a first action; errors explain and offer recovery; preserve user input.
rule/cover-reachable-states, rule/empty-state-action, rule/error-states-recovery, rule/preserve-user-input. See references/surfaces.md.
- Resilience. Require that overflow, extreme data, localization and RTL, and network-failure states be designed; every fetch lands in a designed state.
rule/cover-reachable-states. See references/surfaces.md. Whether the built UI renders them correctly is ui-audit's check.
- Accessibility as a product concern. Every control has an accessible name; the primary flow is completable by keyboard with visible focus; state and consequence are understandable, not just labeled.
rule/accessible-name-required, rule/keyboard-complete-flow, rule/no-custom-focus-bypass. Route axe-style markup checks to ui-audit. See references/interface-quality.md.
Review output
In review and harden modes, lead with findings ordered by user impact (P0-P3), each with location, verification status, rule ID, user consequence, and the smallest concrete fix with the skill that owns it. Keep findings at decision altitude; a line-level code or framework fix is ui-audit's output. Full severity rubric and finding format in references/interface-quality.md > Severity rubric.
Linters vs agent guidance
Deterministic, structural, single-file checks (control selection by option count, nested modals, missing accessible names) belong in the consuming project's linter, wired to that project's components; judgment that needs product context (which object, what consequence) stays here. See references/lint-patterns.md for the decision tree and the rules worth encoding.
Gotchas
- A mockup with only a happy path is a P1 minimum, not a pass (
rule/cover-reachable-states).
- A select with two options is not a style nit; it is
rule/control-matches-cardinality, and a project linter can catch it (see references/lint-patterns.md).
Related skills
ui-design: visual direction and building the decided interaction in code.ui-audit: the built result, both rendered quality and accessibility-markup audit and React or Next diff-level UX bug hunt with a ship verdict.copywriting: exact wording for names, errors, and empty and loading copy; defines shared copy rule IDs in its references/ui-states.md.typography-audit, ui-animation: type and motion.
1---2name: product-design3description: Decides what an interface should do before UI is built or audited: interaction choice, action scope and consequence, reachable states, resilience, and accessibility as task completion. Works from a brief, spec, mockup, intent, or existing UI. Use when asked "is this the right interaction", "design the flow", "what control should this use", "what should this action affect", "which states should this have", "make this resilient", or "what breaks here". For building or styling use ui-design; for built-code audits use ui-audit; for copy wording use copywriting.4---56# Product Design78Decide what the interface should do, then route who builds and verifies it: pick the right interaction, make scope and consequence clear, cover reality beyond the happy path. This skill owns the decision; it routes the build, verification, and copy out (ownership map in Related skills).910- **IS:** the decision layer. From a brief, spec, mockup, intent, or existing UI: choose the right interaction and control, name the object, scope, and consequence of actions, enumerate every reachable state, set resilience expectations, require accessibility as task completion. It decides, then routes build and verification out.11- **IS NOT:**12 - building or styling UI, visual direction, palettes, type: use `ui-design`.13 - auditing the built result (rendered quality, a11y markup, keyboard, layout, performance, type surface, React/Next code-level UX with a ship verdict): use `ui-audit`.14 - copy wording, persuasion, or AI-ism removal: use `copywriting`.15 - deep typography or motion: use `typography-audit` or `ui-animation`.1617## product-design or ui-audit?1819Dispatch on the artifact, not the topic. Both care about states and interaction; they act at different moments.2021| You have... | The question is | Use |22|---|---|---|23| A brief, spec, mockup, intent, or a UI you decide *about* | What should exist: the right interaction, the action's name, which states *should* be reachable | `product-design` |24| Code, a diff, or a running UI you decide *on* | Is the built result right: states covered, accessible, renders and behaves correctly, ready to ship | `ui-audit` |2526One artifact often needs both in sequence: `product-design` decides the states that must exist, `ui-audit` verifies the built code and rendered result implement them. Looking at code or a running screen is `ui-audit`'s turn. This skill reviews the *decision* and stops at decision altitude; it never writes line-level code fixes.2728## Operating contract2930- Cite a stable rule ID for every finding or non-mechanical decision. Never invent an ID; if none fits, record a coverage gap.31- The project's design system and `AGENTS.md` outrank this skill's defaults. Defer to them.32- Never restyle or rebuild. Decide, then route the build to `ui-design`.33- One mode per request, resolved from the user's verb before acting.3435## Request modes3637Resolve the mode from the user's verb and artifact, then load only that mode's references.3839| Mode | Dispatch when the user asks for | Load |40|------|--------------------------------|------|41| **shape** (default) | "design the flow for", "what control here", "how should this work", "is this the right pattern", a brief with no settled UI | `references/product-judgment.md`, `references/surfaces.md` |42| **spec** | "spec the right interaction", "define the expected states", judgment applied before or during a build | `references/surfaces.md`, `references/naming-and-copy.md`; route the build to `ui-design` |43| **review** | "review this for product correctness", "what's wrong with this UX decision", "audit this flow" | `references/interface-quality.md`, `references/rules.md` |44| **action** | "what should this action affect", "which object or scope does this action cover", or action reversibility is unsettled | `references/naming-and-copy.md`; route final wording polish to `copywriting` |45| **harden** | "make this resilient", "what breaks here", error, permission, offline, and destructive paths | `references/surfaces.md`, `references/interface-quality.md` |4647Modes chain: shape leads into spec; review leads into harden. When intent is ambiguous, use the narrowest mode the verb supports. A URL, screenshot, route, or component identifies scope; it does not authorize edits.4849A material decision: see `references/product-judgment.md`.5051## Decision authority5253Conflict order, highest first:54551. The user's explicit goal and constraints.562. Verified user and product evidence, and what the system actually does.573. Project-canonical guidance: `AGENTS.md` or `CLAUDE.md`, the project's design system, routed sibling skills.584. Sibling-skill ownership: route, do not duplicate (ownership map in Related skills).595. This skill's product design standards (below).606. General interface and platform conventions.6162When a request spans authorities, name the owning skill and hand off.6364## Workflow6566```67Product design pass:68- [ ] Step 1: Classify the request into one mode69- [ ] Step 2: Locate authority (user constraints, project design system, AGENTS.md)70- [ ] Step 3: Load only that mode's reference files71- [ ] Step 4: Identify object, scope, and consequence of each action in scope72- [ ] Step 5: Enumerate reachable states; check coverage (surfaces.md)73- [ ] Step 6: Apply standards; cite a stable rule ID per finding or decision74- [ ] Step 7: Emit output (review and harden use P0-P3); route follow-on work to siblings75```7677For shape, spec, harden, or any material product or flow change, write the compact internal brief in `references/product-judgment.md` before proposing UI: user, job, current behavior, desired outcome, success signal, non-goals, object, action, consequence, reversibility, permissions, and open decisions.7879## Product design standards8081Five pillars. Each cites its governing rule IDs in `references/rules.md` and its reference.8283- **Right interaction.** Pick the control from the choice's shape; keep options visible and reversible; prefer inline disclosure over a modal; choose the smallest coherent intervention. `rule/control-matches-cardinality`, `rule/navigation-vs-action`, `rule/inline-before-modal`, `rule/smallest-intervention`. See `references/product-judgment.md`.84- **Action naming.** Name the object, scope, and consequence; destructive CTAs use Verb plus Noun, never "Confirm" or "OK"; make friction proportional to impact and offer undo when honest. `rule/name-object-scope-consequence`, `rule/destructive-names-action`, `rule/destructive-proportional`. See `references/naming-and-copy.md`.85- **State coverage.** Design every reachable state, not just the populated one; empty states name the object and a first action; errors explain and offer recovery; preserve user input. `rule/cover-reachable-states`, `rule/empty-state-action`, `rule/error-states-recovery`, `rule/preserve-user-input`. See `references/surfaces.md`.86- **Resilience.** Require that overflow, extreme data, localization and RTL, and network-failure states be designed; every fetch lands in a designed state. `rule/cover-reachable-states`. See `references/surfaces.md`. Whether the built UI renders them correctly is `ui-audit`'s check.87- **Accessibility as a product concern.** Every control has an accessible name; the primary flow is completable by keyboard with visible focus; state and consequence are understandable, not just labeled. `rule/accessible-name-required`, `rule/keyboard-complete-flow`, `rule/no-custom-focus-bypass`. Route axe-style markup checks to `ui-audit`. See `references/interface-quality.md`.8889## Review output9091In review and harden modes, lead with findings ordered by user impact (P0-P3), each with location, verification status, rule ID, user consequence, and the smallest concrete fix with the skill that owns it. Keep findings at decision altitude; a line-level code or framework fix is `ui-audit`'s output. Full severity rubric and finding format in `references/interface-quality.md` > Severity rubric.9293## Linters vs agent guidance9495Deterministic, structural, single-file checks (control selection by option count, nested modals, missing accessible names) belong in the consuming project's linter, wired to that project's components; judgment that needs product context (which object, what consequence) stays here. See `references/lint-patterns.md` for the decision tree and the rules worth encoding.9697## Gotchas9899- A mockup with only a happy path is a P1 minimum, not a pass (`rule/cover-reachable-states`).100- A select with two options is not a style nit; it is `rule/control-matches-cardinality`, and a project linter can catch it (see `references/lint-patterns.md`).101102## Related skills103104- `ui-design`: visual direction and building the decided interaction in code.105- `ui-audit`: the built result, both rendered quality and accessibility-markup audit and React or Next diff-level UX bug hunt with a ship verdict.106- `copywriting`: exact wording for names, errors, and empty and loading copy; defines shared copy rule IDs in its `references/ui-states.md`.107- `typography-audit`, `ui-animation`: type and motion.
HK-hub