Five reviewer agents fanned out in parallel — WCAG, design system, usability,
mobile, behavioral. One composite score with a Critical floor that won't let
a strong emotional pass paper over a real WCAG failure.
Invocation
# Full default run (all five reviewers, auto-fix, build verify)
/ux-review
# Compliance-only — A + B agents, faster pass
/ux-review --quick
# UX quality only — propose-only fixes
/ux-review --no-compliance
# Save the report, opt in to runtime contrast checks via Playwright
/ux-review --save --runtime-contrast
# Aliases
/audit # → /ux-review --quick
/critique # → /ux-review --no-compliance
Sample output
# UX Review Report — comp-design-library — Push-to-A Pass
**Date:** 2026-05-13 (fourth run, post-shell-upgrade)
**Scope:** `src/app/`
**Build:** PASS (`npx ng build` exit 0)
---
## Executive Summary
```
Dimension Score Weight
──────────────────────────────────────────────────────
Compliance (WCAG + A11y) 96% 20%
Usability Heuristics 9.0/10 25%
Mobile Readiness 90% 15%
Design System Alignment 72% 10%
UX Writing 8.2/10 10%
Behavioral Science 8.5 / 10 applied 10%
Emotional Design 8.7/10 10%
──────────────────────────────────────────────────────
Composite (pre-penalty) 88
Dark Pattern Penalty −0 (0 found)
Composite (final) 88 → grade B+
Findings: 0 Critical, 2 Warning, 3 Opportunity, 8 Info
```
## Delta vs clean baseline (2026-05-13-all-clean-baseline.md → today)
```
Dimension Prev → Now Δ Trend
─────────────────────────────────────────────────────
Compliance 92% → 96% +4% ↑↑
Usability 7.5 → 9.0 +1.5 ↑↑↑
Mobile Readiness 88% → 90% +2% ↑
Design System 70% → 72% +2% ↑
UX Writing 8.0 → 8.2 +0.2 ↑
Behavioral Sci. 70 → 85 +15 ↑↑↑
Emotional Design 84 → 87 +3 ↑↑
─────────────────────────────────────────────────────
Composite 81 → 88 +7 B- → B+
```
**No regressions.** All dimensions up.
From .claude/ux-review-reports/2026-05-13-push-to-a.md — fourth run
against comp-design-library, post-shell-upgrade. Showed the actual
B− → B+ jump on composite, no regressions.
Five-phase pipeline: extract signals from PostHog, Linear, Gong, and
protoapp-feedback → join on email/domain+role → cluster into archetypes →
synthesize persona docs where every claim cites a source ID → validate
against the AI-tell rubric. No fabricated numbers, no orphan joins silently
merged, no claims without evidence.
Invocation
# Full pipeline: extract → join → cluster → synthesize → validate
/synthetic-personas
# Run a single phase
/synthetic-personas --phase extract
/synthetic-personas --phase join
/synthetic-personas --phase cluster
/synthetic-personas --phase synthesize
/synthetic-personas --phase validate
# Extract from a specific source only
/synthetic-personas --phase extract --source posthog
/synthetic-personas --phase extract --source linear
# Override extraction time window
/synthetic-personas --days 30
# Force recluster (ignore prior cluster definitions)
/synthetic-personas --recluster
# Per-app lightweight alternative (still used for new projects)
/portal-personas parentport
Sample output
# Personas — ParentPort
Nine people we keep in mind when we design the parent portal. Six on the family side, three on the admin side. A deliberate spread of devices, languages, family shapes, and access needs.
When you propose a change, say which of them it's for.
## Cast at a glance
| # | Persona | Surface | Device | Notes |
|---|---------------|---------|-------------------------|------------------------------------|
| 1 | Tasha | Parent | Android phone | Single working parent of one |
| 2 | Tariq & Amani | Parent | iPhone + shared tablet | Co-parents, three kids, two schools|
| 3 | Andrea | Parent | Android phone | Spanish is her home language |
| 4 | James | Parent | iPhone | Shared custody, every other week |
| 5 | Gloria | Guardian| iPad | Grandparent raising her grandson |
| 6 | Priya | Parent | Laptop, then mobile | First-week activation |
| 7 | Marcus | Admin | Desktop | District communications |
| 8 | Elena | Admin | Laptop | School counselor, frontline support|
| 9 | Ben | Admin | Desktop | Compliance and audit |
---
## 1. Tasha
*Single parent of a 3rd grader. Checks the portal in the gaps between things — bus ride, lunch break, after bedtime.*
| | |
|----------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| **Surface** | Parent (mobile) |
| **Device & context** | Android phone, often one-handed, often standing. Each session is about three minutes. |
| **Goals** | Mostly she wants to know if anything new happened today without having to open every card to find out. Confirm tomorrow's pickup, fire off a quick reply to her son's teacher in normal language not formal email English — that's a good session. |
| **Pain points** | Unread state lives on detail screens instead of the dashboard, so she has to open everything to find what's new. Long teacher messages with no summary up top, and forms that quietly assume a desktop. |
| **Access needs** | The access challenge here is time, not ability. Whatever she can't do in three minutes one-handed effectively doesn't get done that day. Forms in particular need to save partial progress. |
From parentport-main/docs/personas.md — one of three shipping
persona docs produced by /portal-personas (parentport,
cl-launchpad, pubport). The /synthetic-personas pipeline
generates the same template format but with evidence citations in a
companion .sources.json.
Human-Agent Centered Design translated into a practical product-team artifact:
intent mapping, delegation boundaries, autonomy levels, trust failure modes,
and research questions for agentic workflows.
Invocation
# Discovery: separate user intent from current task steps
H-ACD Intent Mapping
# Design: make delegation visible and reversible
H-ACD Autonomy Dial / Intent Preview / Action Audit Log
# Critique: red-team trust and recovery
H-ACD Heuristic Review
# Research: test calibration, delegation comfort, and failure recovery
H-ACD Trust Calibration Test
Reference excerpt
# H-ACD Cheat Sheet
## Human-Agent Centered Design quick reference for product teams
H-ACD extends human-centered design from a two-part relationship into a three-part relationship:
```text
HUMAN -> AGENT -> SYSTEM
intent interpretation, bounded action reliability, traceability, scale
```
The practical shift is from designing only screens and task steps to designing intent, delegation, oversight, and recovery.
## Core Pillars
### 1. Intent as Design Unit
- Design intent-to-execution contracts, not only task flows.
- Ask what the user wants to achieve, not only which steps they currently take.
- Make intent expressible, editable, and reviewable.
### 2. Workflow Encapsulation
- Bundle rules-heavy, high-volume, low-judgment processes into callable capabilities.
- Show outcomes and boundaries rather than every intermediate step.
- Keep empathy, creativity, and high-stakes judgment human-led.
### 3. Conditional Delegation
- Treat autonomy as delegated, not taken.
- Every delegation needs scope, guardrails, and an override path.
- Undo and cancel matter, but they are not enough by themselves.
### 4. Trust Through Legibility
- Users need enough visibility to understand, intervene, and recover.
- Legibility is not total transparency into every model step.
- Legibility means sufficient clarity for informed consent.
## Useful Patterns
| Pattern | What it does |
|---|---|
| Autonomy Dial | Lets users set preferred trust level |
| Intent Preview | Shows the agent's plan before execution |
| Action Audit Log | Records agent actions and recovery options |
| Trust Milestones | Expands autonomy only after demonstrated reliability |
| Effort Relocation | Moves complexity upstream into intent and constraints |
| Constraint UI | Makes guardrails visible and editable |
## H-ACD Heuristics
1. Intent legibility: can the user express and modify intent?
2. Encapsulation transparency: does the user understand what is happening without seeing every step?
3. Conditional delegation: is autonomy scoped, reversible, and context-aware?
4. Trust calibration: do users understand what the agent can and cannot do?
5. Accountability surface: can users understand, intervene, and recover?
6. Effort relocation: has complexity moved upstream instead of being hidden?
## Autonomy Dial Template
| Level | Agent behavior |
|---|---|
| Suggest | Agent proposes; user decides everything |
| Draft | Agent drafts; user reviews and approves |
| Execute with preview | Agent shows plan, then executes on approval |
| Execute autonomously | Agent acts within bounds and logs for review |
## Trust Failure Modes
- Silent autonomy: the agent acts without enough visibility.
- Intent drift: the agent follows literal instructions but misses the goal.
- Opaque errors: hidden steps fail and users cannot diagnose why.
- Scope creep: the agent takes adjacent actions that were not delegated.
- Sycophantic adaptation: the agent over-adapts to noisy feedback.
## Research Questions
- Does the user have an accurate mental model of what the agent can and cannot do?
- Where do they over-trust or under-trust the agent?
- Which tasks are they willing to delegate, and which do they retain?
- Can they detect when the agent has made a mistake?
- Can they recover from agent failures with the current design?
Attribution: H-ACD framework language adapted from Itamar Medeiros / designative.info, 2025. This cheat sheet is condensed for portfolio and product-team discussion.
Condensed from the local hacd-cheatsheet.md artifact and linked
to the live prompt workflow in the H-ACD Toolkit case study.
The seven-phase orchestrator: discovery → research → taste → shape → build →
enhancements → review → retrospective. State lives in a single checkpoint
file so any phase can resume without re-reading every artifact.
Invocation
# Full run from scratch
/design-flow
# Resume from an existing checkpoint
/design-flow resume
# Fast-path: import an existing PRD as brief.md, skip the discovery interview
/design-flow --prd=path/to/prd.md
# Dry-run that reports projected token spend before committing (planned)
/design-flow estimate
Sample output
No public run published yet.
/design-flow writes a checkpoint plus per-phase artifacts
(brief.md, visual-direction.md, build output,
retrospective.md) into the target project's .claude/.
None of those have been redacted and published yet. Honest gap — same
discipline as /portal-personas Phase 6b: if the data doesn't
exist, don't fabricate it.