Taste Board PRD — Visual Intelligence System

1. What Is This?

An internal knowledge base that turns visual inspiration into reusable design intelligence. Not a gallery. Not a mood board. A work tool that makes every project faster because the team has already studied what works, why it works, and when to apply it.

Scout → Critique → Extract → Apply → Evolve

2. The Problem

Every new project starts cold. The team rebuilds taste from zero instead of compounding knowledge across projects.

Assaf saves inspiration to mymind (421 entries) and Notion (890 entries) — scattered, unprocessed
When a project starts, everyone googles from scratch
No shared language for "why this works"
Past CE work isn't systematically reviewed for learnings
Visual standards exist but aren't connected to reference examples
Knowledge lives in Assaf's head, not in the system

3. Users & Use Cases

User	Role	Key Action
Assaf	Curator-in-chief	Saves, qualifies, approves entries
Jessica	Art direction	Builds visual theses from pattern clusters
Tatiana	Visual execution	Queries for style references
Ogilvy	Copy	Searches voice/tone references
Anton	Quality gate	Compares deliverables against references
Gerri	Coordination	Attaches refs to project briefs
Julia	Research	Adds trend entries, spots patterns

JESSICA / TATIANA → Project Kickoff

"We're starting visual identity for a fintech. Show me everything tagged fintech + premium + minimal."

System returns matched entries with extracted principles + suggested visual thesis.

OGILVY → Copy Reference

"Client needs confident B2B voice. Show me entries where copy critique mentions 'authority without arrogance'."

System returns entries with copy extractions matching the voice pattern.

ANTON → Quality Check

"Compare this deliverable against our approved references for editorial layout."

System returns reference-grade entries for benchmarking.

JULIA → Trend Scouting

"Add 5 new entries from this week's design trends. Extract what's emerging vs. noise."

New entries enter queue with auto-extraction, cross-referenced against existing patterns.

GERRI → Brief Assembly

"New project brief for restaurant branding. Attach relevant references."

Queries by industry + type, attaches top-scored entries to the brief.

KITT → Standards Evolution

"This approved output should become a reference. Add to knowledge base."

Internal work enters taste board, proven patterns graduate to STANDARDS.md.

4. System Architecture

4.1 High-Level Flow

Inputs

mymind

Assaf's saves
421 entries

Notion DB

Team inspo library
890 entries

Manual Add

URL paste / Discord
/taste add [url]

CE Internal

Approved project
work loops back in

▼

Step 1 — Ingest

Capture

Screenshot · Metadata · Tags · Source URL

▼

Step 2 — AI Extraction

The Intelligence Layer

"What works" · "Why" · "Apply when" · "Avoid when" · Copy critique · Visual principles

▼

Step 3 — Human Qualification

Assaf / Jessica Review

Approve / reject / refine · Score 1-10 · Add context AI missed

▼

Knowledge Base

Structured entries — searchable by tag, principle, style, industry, copy pattern

▼

Human

Web UI

Browse, search, filter, qualify, review queue

Agents

Query API

Tag/principle search → entries + patterns + thesis

▼

Project Briefs

Gerri attaches refs

Visual Thesis

Jessica builds from clusters

Quality Gate

Anton benchmarks output

Copy Refs

Ogilvy pulls voice patterns

Trend Radar

Julia spots emerging patterns

Standards

Proven patterns graduate to STANDARDS.md

4.2 Agent Integration — Project Workflow

📋 New Project Arrives

▼

Gerri

Receives brief → queries taste board
by industry + style + context

▼

🔍 Taste Board Query

Returns 5-8 relevant refs
with extracted principles

▼

Jessica

Builds visual thesis from patterns

Tatiana

Uses refs for execution style

Ogilvy

Uses copy patterns for voice

▼

🐀 Anton

Compares output vs. references
Pass → Kitt | Fail → back to agent

▼

Kitt → Assaf

Final review & approval

▼

🔄 Approved work re-enters Taste Board

The compounding loop — every project makes the next one better

5. Data Model

{
  // ─── CAPTURE LAYER ───
  "id": "uuid",
  "source": "mymind | notion | manual | internal",
  "title": "Stripe Dashboard Redesign",
  "url": "https://stripe.com/dashboard",
  "screenshot": "screenshots/stripe-dash.webp",
  "tags": ["fintech", "dashboard", "data-viz"],
  "type": "website | brand | editorial | packaging",
  "industry": "fintech",

  // ─── EXTRACTION LAYER (AI + human-verified) ───
  "extraction": {
    "visual": {
      "what_works": "Dense data + generous whitespace + muted palette",
      "why": "Reduces cognitive load — focus on numbers, not chrome",
      "techniques": ["whitespace-as-hierarchy", "type-scale-contrast"],
      "apply_when": "Data-heavy interfaces where trust matters",
      "avoid_when": "Consumer/playful — feels too clinical"
    },
    "copy": {
      "what_works": "Action labels are verbs, not nouns",
      "why": "Reduces decision time — user knows what clicking does",
      "voice_pattern": "confident-minimal"
    },
    "overall_principle": "Trust through restraint"
  },

  // ─── QUALIFICATION LAYER (human) ───
  "qualification": {
    "score": 9,
    "status": "approved | pending | rejected",
    "qualified_by": "assaf",
    "notes": "Reference-grade for any fintech project"
  },

  // ─── INTERNAL WORK (CE projects only) ───
  "internal": {
    "project": "etoro-sow",
    "version": "v2-approved",
    "what_we_learned": "Client pushed back on bold type — approved after seeing in context"
  },

  // ─── CONNECTIONS ───
  "used_in_projects": ["etoro-sow", "flowrx-brand"],
  "graduated_to_standards": false
}

6. The Extraction Engine

This is the core differentiator. Raw inspiration is noise. Extracted principles are signal.

New Entry Added

URL + screenshot + tags

▼

Step 1 — Auto-Extraction (AI)

Tatiana (visual) + Ogilvy (copy) via Gerri

Visual analysis → what works, why, techniques
Copy analysis → voice pattern, messaging strategy
Pattern tags → reusable technique labels
"Apply when / Avoid when" → context guidance
Related entries → connections to existing knowledge

▼

Step 2 — Human Review Queue

Assaf or Jessica reviews

Is AI extraction accurate?
Score 1–10
Add/edit principles the AI missed
Approve · Reject · Refine

▼

Step 3 — Live in Knowledge Base

Only approved entries are queryable

Rejected → archived · Pending → visible but flagged

7. The Feedback Loop

The system compounds. Every project makes the next one better.

📚

Taste Board

Reference library

→

📋

Project Brief

Refs attached

→

🎨

Execution

Informed by refs

→

✅

Approved

By Assaf

→

📏

Standards

Patterns graduate

◄──── Approved work + proven patterns re-enter as references ────►

Graduation Rules

Pattern used successfully in 2+ projects → candidate for STANDARDS.md
Assaf explicitly flags output as "reference-grade" → enters taste board
Daily EOD review (6pm): "Anything approved today that should become a standard?"

8. Build Phases

PHASE 1

Foundation

~15 hours · Week 1

Finalize data model from PRD
AI extraction pipeline (Tatiana + Ogilvy)
Static site with search + filter + review queue
Import 421 mymind entries with auto-extraction
Basic agent query (JSON file read)

PHASE 2

Integration

~10 hours · Week 2

Notion sync cron
Discord /taste add command
Gerri auto-attaches refs to briefs
Usage tracking (which refs → which projects)

PHASE 3

Intelligence

~8 hours · Week 3

Pattern clustering (auto-detect shared principles)
Visual thesis generation
Internal work pipeline
Standards graduation tracking

ONGOING

Refinement

Continuous

Semantic search (beyond tags)
Trend detection (Julia cross-refs)
Quality scoring recalibration
Performance metrics review

Total estimated effort: ~33 hours across 3 weeks. No backend servers — JSON files + static site + Vercel.

9. Migration — What We Already Have

Asset	Count	Status	Migration
mymind entries	421	JSON ready	Import → add extraction
Notion Inspiration DB	890	API accessible	Sync → dedupe → extract
Screenshots	~2,645	On disk	Map to entries, QC
Notion covers	238	Downloaded	Use where better quality
VISUAL-STANDARDS.md	1	Active	Graduation seed
COPY-STANDARDS.md	1	Active	Copy pattern seed

10. Success Metrics

Metric	Target	How
Time to first visual reference	< 2 min	Agent query logs
Briefs with attached references	> 80%	Gerri tracking
Entries with extraction	> 90%	DB stats
References used in projects	> 50/month	Usage logs
Standards graduated from board	2+/month	Git history

11. Risks & Mitigations

Risk	Impact	Mitigation
Review queue bottleneck	Entries pile up	Batch review UI + auto-approve for trusted high-score
AI extraction quality	Bad principles → bad work	Human review gate on everything
Team doesn't query it	Dead gallery	Embed in workflow — Gerri MUST attach refs
Scope creep to full CMS	3-month project	This PRD is the ceiling. JSON + static site only.

12. Open Questions

Review cadence — Daily batch review, or as-they-come? How many minutes/day for qualification?

Score threshold — What score makes an entry "reference-grade" vs. "interesting but not directive"?

Internal work — All approved CE work auto-enters, or only work you explicitly flag?

Team write access — Can anyone add entries, or only Assaf + Jessica + Julia?

Copy vs. visual priority — Equal extraction weight, or visual-first?