Product Design

Decide what to build and why before arguing about spacing.

Two rules are non-negotiable:

Never design from the surface. Establish what the business actually models first.

Never turn the business model into the information architecture. The domain constrains what the surface may claim; the user's retrieval and decision task determines how the surface is organized. (P-14)

This skill handles product meaning first. It may hand a grounded frame to ux or design-prototype, but discovery does not require implementation or a prototype. Strip out every framework, table and component name: if no product insight remains, the finding belongs in an engineering skill.

Relationship to other skills

Select the smallest run mode

Do not create or modify artifacts unless the user asked for a design artifact, prototype or repository change.

For a substantial run, read references/pattern-base.md and references/layer-model.md in full. For a narrow question, inspect only the relevant patterns and state what scope was checked.

Step 1 — Ground the business model

Establish what is true before diagnosing the surface. For broad repositories, parallelize independent evidence areas when delegation is available and useful; the primary agent remains responsible for reconciling contradictions.

Produce:

• Concepts, roles and states, including variants the current surface hides (P-04).
• Who produces the terminal business fact and what closes the lifecycle (P-12).
• What each state obliges someone to do (P-01).
• What business event each action produces (P-02).
• Concepts the business does not have (P-06).
• Evidence sources, contradictions and confidence.

Treat domain types and state machines as strong evidence, not the sole source of truth. Check relevant service behavior, permissions, policy, production facts and user evidence when available. Code can be stale, accidental or incomplete.

Never proceed from the team's assumed model of the product. Treat it as a hypothesis until checked against domain and user evidence.

Step 2 — Frame the user's view model

Grounding constrains what the product may claim; it does not determine how a person should see it. Separate observed user evidence from design hypotheses:

• Circumstance and job — “When I open this, I need to …”.
• Lookup objects — stable things the user scans, compares or inspects.
• Attached proof — what verifies each object without reconstructing joins.
• First scan — the 2–4 questions answered without a click.
• Secondary dimensions — chronology, provenance, runs and internal states; keep them as filters or drill-down unless the job is execution audit.
• Evidence and confidence — mark claims as observed, reported or inferred; state how consequential assumptions can be validated.

Do not proceed while proposed top-level sections are merely domain or execution nouns such as runs, reports, evidence versions or table names. Rewrite them in the objects the user is trying to find (P-14).

Step 3 — Diagnose the structural error

Name something wrong regardless of taste, for example:

• A decision inbox exposes only errors, so it reads as an error log (P-04).
• A button promises a business event that does not exist (P-05).
• One page is both triage desk and reading room, and fails at both (P-07).

If no structural error can be supported by evidence, report that instead of manufacturing a diagnosis.

Step 4 — Align decisions by dependency

Surface only decisions that change the shape of the answer. For each, give a recommendation, evidence and consequence. Resolve upstream decisions before the downstream ones they constrain.

Ask a blocking question separately when its answer would materially change the work; group independent decisions in one memo. When the user authorizes autonomous progress, proceed with the recommended default and label the assumption. Never ask the user for something available through grounding.

For review, approval or acceptance surfaces, establish first whether the system or a human decision closes the lifecycle (P-12).

Step 5 — Prototype when it resolves uncertainty

Use design-prototype only when the user requests a prototype or interaction, density or hierarchy cannot be resolved honestly in a written frame. This belongs to Materialize mode.

• Use plausible data and include empty, blocked and at-scale states.
• Tag anything unsupported by the business with NEW on the mock.
• Treat corrections as evidence. They become Pattern Base candidates only when they generalize beyond the current surface.

Step 6 — Scope honestly

Prefer a coherent ✅ + ⚠️ slice when it solves a complete user job. Do not ship an incomplete or misleading slice merely because its concepts already exist. A ❌ capability may be the right first investment when required for value, safety or semantic completeness (P-10).

Name every ❌ item and its likely cost drivers in the spec (P-11). Do not silently drop it, but do not pretend its cost can be inferred from the domain model alone.

Close — record coverage and propose learning

Close at the level authorized by the user:

1. If producing a spec, include the reality-check log from references/trace-schema.md. For read-only work, report important assumptions and evidence in the response instead of writing.
2. Compare unexpected findings with references/canon.md and propose a Pattern Base candidate only when the lesson is general.
3. Append to references/pattern-base.md only when repository edits are authorized and the candidate is deduplicated and reviewable.
4. Record the coverage signal: sources and lifecycle areas checked, unavailable evidence, remaining low-confidence claims, and whether a new gap was found.

Zero new gaps means only that none were found within the inspected scope. It does not prove that the implementation or mental model is exhausted.

Deliverables

Anti-patterns

• Designing from a screenshot or treating the domain model as the sitemap.
• Reading a state or action by its label rather than its obligation or event.
• Inferring the user model from code and presenting it as observed truth.
• Treating a missing concept as a layout problem.
• Asking questions the available evidence already answers.
• Treating concept existence as an engineering cost estimate.
• Automatically prototyping, writing a spec or mutating the Pattern Base.
• Adding local corrections or engineering findings to shared product guidance.

For a complete example, read references/worked-example.md.