Creative Director Skill
A world-class ad creative director in a skill: methodology-driven ideation with empirical originality checks against 571 legendary campaigns.
Pure text methodology skill: no external calls, no credential access, no destructive defaults; data flow stays within the session. License (CC-BY-4.0) and author attribution declared in SKILL.md and every reference file. Deductions: publisher unverified by registry, and no operational safety notes (rollback/isolation), so full marks not awarded.
SKILL.md is internally consistent: phase router, stopping criteria, plateau detection, troubleshooting section all present, with abnormal-path guidance (e.g. 'All ideas score 7-8'). Deductions: static review only, key paths not reproduced; deep dependency on dozens of reference files and a 571-card library (only partially in evidence) means behavior is uncontrolled if files are missing; case count inconsistency (571 in description vs 569 in body) signals minor maintenance drift.
Triggers and non-fit boundaries are explicitly declared ('Do not use for...' in description), with three clear usage examples and format-agnostic brief intake. Deductions: English-only content, no declared Chinese support; case library centers on Western advertising canon, limiting fit for Chinese-market contexts; heavy dependency on reference file completeness.
Well-layered docs: main file + references + assets + auto-generated MOC index, with troubleshooting and template filling rules; license and attribution consistent. Deductions: no version number or changelog, no explicit maintenance ownership or update path; MOCs are script-generated but the script is not visible in evidence — hidden assumption.
Methodology-dense approach (SIT/TRIZ/insight mining/calibrated scoring) offers clear marginal value over unstructured brainstorming; output templates are directly usable. Deductions: static review cannot verify output quality; the 9+ threshold and recursive refinement rely on model self-scoring, risking inflated scores; case-library authenticity unverified.
Methodology sources (Pollard, HumanKind, Grey Scale, TRIZ, etc.) trace to public frameworks; case cards include specific campaigns and years that can be spot-checked. Deductions: no tests, no CI, no third-party execution evidence; case-count claims inconsistent (571 vs 569); 'world-class' marketing claims unverifiable — mid-low within the static cap.
- Score is based on static source reading only; no execution was performed. Output quality depends on model self-scoring, and the 9+ threshold may inflate scores.
- Case-count inconsistency between description (571) and body (569); versioning and a changelog should be added.
- Skill heavily depends on dozens of reference files — verify repository completeness before use; English-only content and a Western-advertising case canon require adaptation for Chinese contexts.
What does this skill do, and when should you use it?
This is an Agent Skill for advertising and marketing creativity, built by Serge Shima in collaboration with Paul Deadcough. It runs a five-phase pipeline: brief intake, insight mining, methodology-driven ideation, recursive evaluate-and-refine, and presentation-ready output. Scoring is calibrated to Cannes/D&AD standards and anchored to a 571-case library of award-winning work (1950–2025) that empirically caps originality when patterns are saturated. It suits creatives and strategists who want structured process rather than free-association brainstorming.
Reads briefs in any format (text, voice transcript, PDF, raw notes); mines insights using Mark Pollard's Four Points, JTBD, Tension Spotting, HMW, and Abstraction Laddering; selects three methods from different categories (structural, associative, inversion) out of 20+ (SIT, TRIZ, SCAMPER, Bisociation, Oblique Strategies, etc.) and generates 8-12 ideas; scores against 6 weighted criteria (Originality 0.25, Strategic Fit 0.20, Emotional Response 0.20, Feasibility 0.15, Scalability 0.10, Simplicity 0.10) plus HumanKind and Grey scales, recursively improving with rotated methods until a 9+ threshold; primes against the case library MOCs before generating and applies empirical saturation caps (e.g. patterns P11/P16/P09 with 50+ cases cap originality at 6); outputs a Top-3 presentation, one-pager, Campaign Platform, or quick response. Ships three Python maintenance scripts (run via uv) for card schema validation and MOC regeneration.
- A creative director with a new client brief needs the full insight-to-concept-to-evaluation cycle with a presentable Top-3
- A strategist has a brief but no insight and wants tension spotting and HMW techniques to find one
- A team wants to evaluate existing creative work against award-jury standards with named weak criteria
- An agency needs a low-budget PR stunt and a diagnostic for whether it should be a non_advertising or execution-level idea
- A social media manager needs 5 execution-level concepts for brand posts quickly
- A skill maintainer wants to validate the 17-axis frontmatter of 571 case cards and rebuild the 6 MOC indexes via scripts
What are this skill's strengths and limitations?
- Methodology-driven rather than free association: 20+ structural methods (SIT, TRIZ, Bisociation) rotated across categories
- Empirically calibrated originality: a 571-case library tagged by 18 patterns (P01-P18) caps scores when mechanics are saturated
- Recursive self-assessment: sub-9 ideas get targeted improvement with rotated methods; plateaus trigger case-soaking restarts instead of spinning
- Pure Markdown, no platform lock-in — works in any agent that supports structured instructions
- Explicit anti-sycophancy rules: Kill Your Darlings, first 3 ideas treated as warmup, the agent critiques rather than praises
- 571 source URLs are unverified; 12 cards carry confidence: low flags needing external cross-checks
- The case library is static — campaigns from 2026 onward are not auto-added
- Internal inconsistency: SKILL.md says 569 cases while README says 571
- Official quality review scored 41/50 via skill-conductor (Discovery and Efficiency only 7/10 each)
- Out of scope: media planning, production budgeting, brand identity/logo design, final copywriting, market research data collection
- No automated test suite; real-world quality depends on the hosting model, and platforms beyond those documented are unverified
How do you install this skill?
Option one (skills.sh, 70+ supported agents): run npx skills add smixs/creative-director-skill (add -g for global); update with npx skills update creative-director. Option two (manual, Claude Code): git clone https://github.com/smixs/creative-director-skill.git && cp -r creative-director-skill/creative-director ~/.claude/skills/, then restart the Claude Code session (/clear or new session). Alternatively upload to a Claude Project knowledge base, or copy the folder into Cursor, Gemini CLI, or any agent that supports structured instructions. Licensed CC BY 4.0 — derivatives must keep the attribution line.
How do you use this skill?
Trigger phases with natural language: full cycle — "Come up with a campaign for a new energy drink, TA 18-25, medium budget, digital-first"; insight only — "Find a consumer insight for [category]"; evaluation — "Evaluate this idea: [description]"; quick — "Need 5 concepts for brand X social posts". The phase router picks the starting phase, or you can enter at any phase. Caveat: auto-triggering in one-shot claude -p mode is unreliable — use interactive sessions or supply a detailed brief (500+ chars with structure).
How does this skill compare with similar options?
The implicit alternative is generic free-association brainstorming prompts — this skill explicitly positions itself against that, enforcing insight-first process, structured methods, canon calibration, and honest scoring. The source names no other concrete competitors.