AI Skill Report Card
Directing AI Creative Production
Quick Start14 / 15
Given a creative brief, run the five-stage pipeline:
- Ideate — generate 30-50 raw concepts, no filtering
- Adapt — reshape top candidates into target tone/style
- Curate — cut to the 5% that actually work
- Transpose — translate the surviving idea across formats
- Prototype — build a cheap draft to validate before full production
Example invocation: "Give me 40 campaign concepts for a vegan sneaker brand targeting Gen Z skaters, then narrow to the 3 best and sketch one as a storyboard."
Recommendation▾
Add a bad-outcome example (e.g., a poorly curated or over-polished prototype) to contrast with the good examples shown
Workflow14 / 15
Progress:
- [ ] Stage 1: Ideation — produce volume, defeat blank-page fear
- [ ] Stage 2: Style adaptation — fit tone/genre/visual language to audience
- [ ] Stage 3: Curation — ruthlessly cut to the top 5%
- [ ] Stage 4: Cross-media translation — port the surviving idea to new formats
- [ ] Stage 5: Prototyping — produce a cheap, fast draft for hypothesis testing
Stage 1: Ideation
- Target quantity over quality: 30-50 distinct concepts minimum.
- Force variation axes deliberately (emotion, format, metaphor, absurdity level, cultural reference) so outputs aren't variations of one idea.
- No self-censoring at this stage — include deliberately weird/extreme options; they often seed the eventual winner.
- Output as a flat numbered list, one line each, no elaboration yet.
Stage 2: Stylistic adaptation
- Define target audience in concrete terms first (age, platform, cultural context, competitor references).
- Build a short "voice spec": 3-5 adjectives + 2 reference examples (brands/creators) + 1 explicit anti-example (what to avoid).
- Re-render chosen concepts through that voice spec, not from scratch.
- Check: would this be mistaken for a different brand/audience? If yes, tighten the spec.
Stage 3: Curation ("taste")
- Apply explicit cut criteria before judging, e.g.: originality, feasibility, emotional resonance, on-brand fit, memorability.
- Score each surviving concept 1-5 per criterion rather than vague gut reaction — forces articulation of why something is cut.
- Target ratio: keep ~5%. If more than ~10% survive, criteria were too lenient — re-cut.
- Write one sentence per kept item explaining the specific spark that justified survival — this becomes the brief for Stage 4.
Stage 4: Cross-media translation
- Identify the core idea independent of its original medium (strip it to a single sentence) before transposing.
- Map explicitly: source medium → target medium, naming what must change (pacing, composition, rhythm) vs. what must stay invariant (core emotion/message).
- Produce one translated artifact per target format requested (e.g., text brief → image prompt → storyboard beat → audio mood description).
Stage 5: Generative prototyping
- Favor speed and cheapness over polish — goal is a disposable test artifact, not a deliverable.
- Timebox each prototype (minutes, not hours of iteration).
- Attach a one-line hypothesis to each prototype ("this tests whether X resonates") so results are evaluable.
- Stop as soon as the hypothesis is confirmed/denied — don't polish past that point.
Recommendation▾
Example 2 and 3 describe output format rather than showing full concrete output — render the actual table/full output for Example 2 to strengthen concreteness
Examples15 / 20
Example 1: Input: "Ideate: 10 taglines for a solar-powered phone charger." Output:
- "Charged by the sun, unplugged from everything else"
- "Your pocket, powered by a star"
- "Never beg an outlet again"
- "Sunlight in, silence out"
- "The grid is optional now"
- "Burn daylight, not battery anxiety"
- "Off-grid never looked this charged"
- "A tiny sun for your pocket"
- "Power that doesn't cost the planet"
- "Plug into nothing but the sky"
Example 2: Input: "Curate these 20 logo concepts down to the best 1-2, criteria: memorability, scalability, originality." Output: Table scoring each 1-5 per criterion, total score column, top 2 highlighted with one-sentence justification each (e.g., "#14 wins: silhouette reads instantly at favicon size, no competitor uses this negative-space trick").
Example 3: Input: "Translate this idea into three formats: 'A city where silence is currency.'" Output:
- Image prompt: "Hyperdense neon metropolis, citizens whispering, vending machines that dispense noise-cancelling tokens, muted color palette broken by glowing price tags"
- Storyboard beat: "Opening shot: crowded subway, total silence except one coin drop — everyone turns, tense"
- Audio mood: "Near-silence with sudden sharp foley accents; ambient hum replaces traditional score"
Recommendation▾
Include guidance on handling edge cases like client rejecting all curated options or ideation producing only convergent/safe ideas despite the process
Best Practices
- Always separate ideation from judgment — generating and curating in the same pass kills volume and originality.
- Make curation criteria explicit and written down before scoring, not improvised per-item.
- Keep a "graveyard" list of cut concepts; today's reject is tomorrow's seed for a different brief.
- Prototype at the lowest fidelity that still tests the actual hypothesis — don't over-render.
- When adapting tone, always include an anti-example — "sound like X, not like Y" prevents genre drift.
Common Pitfalls
- Judging ideas as they're generated — this collapses quantity and produces safe, convergent output.
- Vague curation ("I just like it") — always force a named criterion.
- Skipping the "core idea" extraction before cross-media translation, resulting in literal/awkward ports (e.g., describing the text instead of transposing the feeling).
- Over-polishing prototypes meant only for hypothesis testing — wastes the cost advantage of the stage.
- Narrowing the voice spec too early in ideation, which reintroduces blank-page fear instead of deferring style to Stage 2.