Bezel Team hitl-screen-design design-001
internal prototype · canonical JSON + Dreamborn Forge HTML
internal generated
hitl-screen-design · supabase_json

Bezel Team hitl-screen-design design-001

The current design process produces useful intent, architecture, and plans, but it does not reliably break the product into human-reviewable screens, wireframes, and states before implementation. That creates a gap where agents can build technically plausible software that the product owner never truly visualized or approved.

Planning Surface

Use this to decide what happens next.

Status

draft

Phase

design-001

problem

The current design process produces useful intent, architecture, and plans, but it does not reliably break the product into human-reviewable screens, wireframes, and states before implementation. That creates a gap where agents can build technically plausible software that the product owner never truly visualized or approved.

goal

Before any production build starts, Bezel must guide the customer through a concrete screen and workflow review: what screens exist, what each screen does, what states it has, what data it needs, what actions are possible, and what the customer is approving.

Agent Handoff
Start Here

The current design process produces useful intent, architecture, and plans, but it does not reliably break the product into human-reviewable screens, wireframes, and states before implementation. That creates a gap where agents can build technically plausible software that the product owner never truly visualized or approved.

Completion Evidence
  • Customer can explain what will be built by pointing to screens, not prose.
  • Zero build tasks are dispatched for screens lacking approved records.
  • Implementation defects caused by misunderstood screen behavior decrease over successive projects.
  • First customer approves screen map and top wireframes before build starts.
Problem

The current design process produces useful intent, architecture, and plans, but it does not reliably break the product into human-reviewable screens, wireframes, and states before implementation. That creates a gap where agents can build technically plausible software that the product owner never truly visualized or approved.

Goal

Before any production build starts, Bezel must guide the customer through a concrete screen and workflow review: what screens exist, what each screen does, what states it has, what data it needs, what actions are possible, and what the customer is approving.

Success Metrics
  • Customer can explain what will be built by pointing to screens, not prose.
  • Zero build tasks are dispatched for screens lacking approved records.
  • Implementation defects caused by misunderstood screen behavior decrease over successive projects.
  • First customer approves screen map and top wireframes before build starts.
Structured Payload

Machine-readable source fields

goal

Before any production build starts, Bezel must guide the customer through a concrete screen and workflow review: what screens exist, what each screen does, what states it has, what data it needs, what actions are possible, and what the customer is approving.

kind

bezel.hitl-screen-design

name

HITL Screen Design Layer

problem

The current design process produces useful intent, architecture, and plans, but it does not reliably break the product into human-reviewable screens, wireframes, and states before implementation. That creates a gap where agents can build technically plausible software that the product owner never truly visualized or approved.

version

1

created at

2026-05-04T12:11:14.124Z

hitl rules
  • Build cannot start from prose-only requirements.
  • Every must-have screen must have an approved screen record.
  • Every must-have flow must have at least one approved flow path.
  • Every production-critical screen must define non-happy-path states before implementation.
  • Open questions on screen purpose, role access, primary action, or data source block build.
  • Customer comments should attach to screen IDs, not free-floating prose.
  • Agents may propose screens and wireframes, but the customer approves or rejects them.
  • Agent recommendations are part of product strategy; building/operating those agents is separate expansion scope unless explicitly included.
  • Every screen review must explain where the customer is in the design/build process and what approval unlocks next.
  • Playgrounds are generated from artifacts for review; they are not source of truth and cannot bypass artifact approval.
project id

bezel-team

updated at

2026-05-04T14:57:54.482Z

next decision

Decide whether wireframes should be rendered as structured HTML inside Bezel first, or generated as image/board artifacts later. Recommendation: start with structured HTML wireframes because they are easier to diff, comment on, and translate into implementation tasks.

success metrics
  • Customer can explain what will be built by pointing to screens, not prose.
  • Zero build tasks are dispatched for screens lacking approved records.
  • Implementation defects caused by misunderstood screen behavior decrease over successive projects.
  • First customer approves screen map and top wireframes before build starts.
wireframe level
v1

Low-to-mid fidelity, intentionally fast: structure, hierarchy, copy blocks, controls, data areas, and states. Not final visual polish.

why

The goal is shared understanding and build correctness, not dribbble-level mockups before validation.

future

High-fidelity visual design can become an optional paid design pass or happen for screens selected as brand-critical.

product principle

For nontechnical product owners, a screen is the unit of understanding. Requirements become real only when the customer can see and approve the screens and flows.

orientation system
principle

The customer should always know where they are in the product-build process and what they are doing next.

action rule

Every review screen must have one primary next action and one clear escape: request changes, ask question, or save for later.

artifact rule

Orientation data should come from phase/gate status and artifact approval state, not static copy.

shell pattern

Persistent project progress rail plus current-step header plus next-action panel.

required questions
  • Where am I?
  • What phase is this?
  • What did Bezel just produce or learn?
  • What decision/action is needed from me?
  • What happens next after I approve or request changes?
  • What is blocked?
required artifacts
purposestatus gateartifact type
Canonical inventory of screens, routes, roles, data needs, actions, dependencies, and priority.Must be approved before wireframes are considered complete.screen-map
Human-reviewable layout and interaction spec for each critical screen.Must be approved before architecture/build tasks are locked.wireframe-spec
User journey map across screens, including alternate and recovery paths.Must be approved before build plan generation.flow-spec
Screen-by-screen state matrix: default, empty, loading, error, success, permission/locked, mobile/responsive where relevant.Must be approved for all screens marked production-critical.screen-state-spec
Identifies where the customer should move from SaaS-shaped screens/manual workflows to AI-first agent workflows, including recommended agents, triggers, human gates, data needs, and whether each agent is included or expansion scope.Must be reviewed before final build plan so customer understands product scope vs optional agent infrastructure scope.ai-first-opportunity-map
Defines where multi-model research runs, which models critique which decisions, and how outputs become accepted artifacts.Must be defined before automating model review gates.model-collaboration-protocol
Defines how artifacts become sandboxed visual playgrounds for HITL review and how feedback flows back into canonical artifacts.Must be reviewed before implementing playground generation.playground-generation-protocol
screen record schema
id

stable screen id, e.g. dashboard.home

name

Human-readable screen name

priority

must | should | later

user roles
  • Who can see/use this screen
build phase

M-XX or TBD

data inputs
  • Data needed to render the screen
data outputs
  • Data created or changed by this screen
integrations
  • External/internal services this screen touches
job to be done

What the user is trying to accomplish here

open questions
  • Questions that block approval
route or entry

URL route, modal, tab, or embedded surface

primary actions
  • Actions the user can take
customer visible flow
stepoutputcustomer experience
Product IntakeIntent summary and open questions.Customer explains the product, users, workflows, business rules, and desired outcome.
Screen InventoryScreen map with purpose, owner/user, route, data, actions, and priority.Bezel proposes every screen needed for the product, grouped by user journey and role.
AI-First Opportunity ReviewOpportunity map with recommended agents, workflow shifts, expected value, and expansion-scope flags.Bezel shows where the proposed product should stop copying SaaS screens and instead use agents, approvals, and automated workflows.
Flow ReviewPrimary flows, alternate flows, and blocked/error paths.Customer reviews how users move through the product from entry to completion.
Wireframe ReviewApproved wireframe set.Customer sees low/mid-fidelity wireframes for each critical screen and can request changes before build.
Playground ReviewPlayground feedback attached to screen/component ids and promoted into artifact revisions.Customer opens an interactive sandbox generated from approved/draft wireframe artifacts to see the product before build.
State ReviewState-complete screen spec.Customer reviews default, empty, loading, error, success, permission, and edge states for screens that need them.
Build Plan ApprovalBuild-ready plan and acceptance criteria.Customer approves what will be built, in what order, and what counts as done.
multi model collaboration
examples
  • Claude critiques whether the flow is understandable and product-focused.
  • Gemini suggests missing screens, alternate states, and AI-first workflow opportunities.
  • Codex/OpenAI checks whether the screen spec is implementable and testable.
principle

Screens, flows, and wireframes should be critiqued by multiple models before customer approval when the decision materially affects product direction.

artifact rule

Store model review evidence and synthesis before build plan approval.

ui implications for bezel
product owner view
  • Screen map board grouped by journey/role
  • Review queue for screens needing approval
  • Wireframe viewer with comments and approve/request changes controls
  • State checklist per screen
  • Change summary between revisions
  • Build readiness indicator based on approved screens/states
  • AI-first opportunity map with Included vs Expansion labels for recommended agents
internal agent view
  • Screen IDs become build task anchors
  • Wireframe/spec links are injected into task briefs
  • Review feedback creates revision tasks instead of ad hoc chat
  • Completion verification checks implemented UI against approved screen-state spec
first implementation slice
name

Screen Map + Wireframe Approval MVP

scope
  • Create screen-map artifact from intake
  • Render screen cards in Bezel Studio project view
  • Generate simple wireframe specs for top 3 must-have screens
  • Support approve/request-changes per screen
  • Block build plan until required screens are approved
  • Create AI-first opportunity map from intake and screen review
non goals
  • Full Figma integration
  • Pixel-perfect visual design
  • Realtime collaborative editing
  • Automatic implementation from wireframes without build artifacts