Created in collaboration with Paul Deadcough.
An AI creative director that generates advertising concepts using world-class methodologies, scores them against Cannes/D&AD-calibrated criteria, recursively refines until the quality threshold is reached, and calibrates every idea against a library of 571 legendary campaigns to detect saturation and ensure originality.
Not a brainstorming toy. A structured creative process that mirrors how top agencies (Droga5, Wieden+Kennedy, Mother) actually work — insight before ideas, methodology over free association, honest evaluation over flattery, calibration against the canon over invention from zero.
What's New in v2.0
- 571-case library with full Insight / Mechanic / Why-it-worked / Steal structure on every card
- 17-axis frontmatter schema — every card tagged across pattern (P01-P18), industry, format, emotion (Tier 1/2/3), budget, idea_type (Pollard 7-level), awards, and more
- 6 MOCs (Maps of Content) for fast lookup by pattern, emotion, format, industry, budget, and chronological index
- 4-point case library integration in workflow — priming before generation, originality empirical cap, RESTART case-soaking, pattern calibration before exit
- 3 maintenance scripts (Python via
uv) — schema validation, MOC generation, graph wikilinks - Pollard 7-level idea taxonomy replacing the previous 3-level system
- 3-tier emotion hierarchy with 30+ specific emotion values
- Activation toolkit with 9 formats and the Non-advertising vs Execution test
- Pre-Mortem template before final delivery
Quality reviewed via skill-conductor: 41/50 production-grade (Discovery 7/10, Clarity 9/10, Efficiency 7/10, Robustness 8/10, Completeness 10/10).
What It Does
Feed it a brief in any format — text, voice transcript, PDF, raw notes — and it runs a full creative cycle:
- INTAKE — extracts the brief's DNA: product, audience, objectives, constraints + classifies the required idea level using Pollard 7-level taxonomy (business / brand / tagline / advertising / campaign / non_advertising / execution)
- INSIGHT — mines consumer insights using 7 proven techniques (Mark Pollard, JTBD, Tension Spotting, HMW, Abstraction Laddering)
- IDEATION — primes against the canon (scans 5-7 cases from the relevant MOC), then generates 8-12 ideas using 3 methods from different categories (structural × associative × disruptive), rotating between 20+ methodologies, with a Tension test on each
- EVALUATE + REFINE — scores against 6 weighted criteria + HumanKind + Grey Scale, then recursively improves; Originality is capped empirically by case-library saturation (3+ canonical cases of the same mechanic → cap at 7); when stuck, RESTARTs through case-soaking (read 8-12 canonical cards, remix allowed); Pattern Calibration + Pre-Mortem before exit
- ARTICULATE — outputs in a presentation-ready format (one-pager, top-3, campaign platform, or quick response)
You can also enter at any phase: jump to insight mining, evaluate an existing idea, or generate concepts from a known insight.
Why This Exists
Most AI "creative" tools generate ideas by free association — producing volume without structure. The result: hundreds of mediocre concepts that nobody can evaluate, often unknowingly recycling work that already exists.
This skill enforces the discipline that separates award-winning work from filler:
- Insight-first — no ideation without a validated consumer tension
- Structural methods — SIT, TRIZ, SCAMPER, Bisociation, Synectics, not "give me 10 ideas"
- Empirical originality — every idea is checked against 571 real campaigns; if 3+ canonical cases use the same mechanic, originality is capped, not subjectively claimed
- Honest scoring — calibrated against real Cannes winners, with anti-inflation rules
- Recursive refinement — weak criteria get targeted improvement using different methods each pass
- Case-soaking on plateau — when ideas plateau, the skill reads 8-12 canonical cards to re-train its sense for what a strong insight feels like; combining/remixing existing patterns is explicitly allowed (this is how Cannes-grade work is built)
- Kill Your Darlings — the skill argues against its own favorite ideas to test their strength
- Pre-Mortem — before delivery, simulates failure and surfaces the most likely failure modes
What's Inside
creative-director/
├── SKILL.md # Core skill — phase router + 5-phase workflow
├── assets/
│ └── output-templates.md # 4 presentation formats
├── scripts/ # Python via uv (PEP 723 inline deps)
│ ├── validate_schema.py # Frontmatter validation against tag-schema
│ ├── generate_mocs.py # Builds 6 MOCs from cards
│ └── generate_links.py # Adds Related sections (idempotent)
└── references/
├── tag-schema.md # Single source of truth — 17-axis frontmatter contract
├── idea-taxonomy.md # Pollard 7-level taxonomy
├── emotion-hierarchy.md # Tier 1/2/3 + 30+ specific emotion values
├── activation-toolkit.md # 9 activation formats + Non-ad vs Execution test
├── legendary-patterns.md # P01-P18 mechanics + Pre-Mortem template
├── methods-catalog.md # 20 creative methodologies as executable cards
├── method-selection-matrix.md # Task → method routing + rotation rules
├── insight-mining.md # 7 insight discovery techniques
├── scoring-calibration.md # Detailed rubrics + calibration anchors
├── creative-constitution.md # 3-layer evaluation system + feedback rules
├── storytelling-frameworks.md # 6 narrative frameworks for advertising
└── legendary-campaigns/ # 571-case library
├── README.md # Library guide
├── MOC-index.md # All 571 cards by year DESC
├── MOC-pattern.md # Grouped by P01–P18 (18 mechanics)
├── MOC-emotion.md # Grouped by emotion_tier (3 → 2 → 1)
├── MOC-format.md # Grouped by category (film/integrated/stunt_pr/...)
├── MOC-industry.md # Grouped by industry vertical
├── MOC-budget.md # Grouped by budget (low/medium/high)
└── cards/ # 571 individual cards (flat namespace)
└── {id}.md
Case Library
571 advertising campaigns from 1950–2025, tagged across 17 axes, with full structural analysis on every card.
Card structure
Each card has YAML frontmatter (17 axes — see references/tag-schema.md) + 4 sections:
- Insight — the human truth the campaign exploits (one sentence)
- Mechanic — what was actually executed
- Why it worked — psychological / cultural / structural reason
- Steal — pattern-level borrow strategy
- Related — auto-generated wikilinks (pattern hub + 2 sibling cards + emotion match)
Tag schema (17 axes)
id, title, brand, agency, year, country, region, industry, pattern[] (P01–P18), category, idea_type (Pollard 7-level), involvement, channel, duration, goal[], budget, emotion[], emotion_tier (1/2/3), insight_domain, media_epoch, awards[], quality_score (HumanKind 1–10), scalability, risk.
18 patterns (P01–P18)
Format as Idea · Enemy or Conflict · Behavior Inversion · Brand as Activist · Cultural Hijack · Limitation as Power · Invisible Brand · Craft as Message · User as Co-Author · Serialization & Ritual · Absurd as Carrier · Social Experiment · Truth Telling · Product as Proof · Benefit Hyperbole · Long-form Drama · Design as Idea · Tech as Canvas
Lookup strategies
# Browse by axis — open the matching MOC
# Each line has inline-context (brand · year · format · emotion · budget · top_award)
# Filter by combination — ripgrep on frontmatter
rg -l "^budget: low$" cards/ | xargs rg -l "^emotion_tier: 3$"
# All P11 cases since 2020
rg -l "P11" cards/ | xargs rg -l "^year: 202"
# All canonical cases for QSR
rg -l "^industry: qsr$" cards/ | xargs rg -l "^quality: canonical$"
Saturation map (originality pressure)
| Pattern | Saturation | Originality cap |
|---|---|---|
| P11 (Absurd as Carrier), P16 (Long-form Drama), P09 (User as Co-Author) | high (>50 cases) | ≤6 unless structurally new variant |
| P02, P10, P12, P14, P18 | medium | ≤7 if 3+ canonical use same mechanic |
| P01, P15 | low | room for novelty |
This is empirical saturation, not subjective novelty.
Methodologies (20+)
| Category | Methods |
|---|---|
| Structural | SIT/Goldenberg Templates, SCAMPER, TRIZ (10 principles), Morphological Analysis |
| **Associatio |