g-eskayo/marvin/creative

Creative

Produce creative output that is genuinely unexpected — not just competent.

The Core Problem

The default creative response is always the most predictable one: the obvious metaphor, the safe structure, the genre cliche, the first idea that comes to mind. That default is almost always wrong for creative work.

This skill forces divergence before convergence.

Process

1. Resist the first idea

Whatever comes to mind immediately — note it, then set it aside. It's the predictable answer. The first idea is usually:

• The genre default
• The safest interpretation of the brief
• What everyone else would also produce

Generate at least 3 radically different directions before choosing.

2. Reframe the brief

Before generating, ask:

• What is the actual goal underneath the stated request?
• What would this look like if the constraints were inverted?
• What unexpected medium, form, or angle could carry this?
• What's the most interesting wrong interpretation?

The best creative work often solves a different problem than the one stated.

3. Apply constraints deliberately

Constraints produce creativity, not freedom. When stuck:

• Add an arbitrary rule (must use no adjectives, must be under 10 words, must be a dialogue)
• Steal the structure from an unrelated domain (write this like a legal brief / a recipe / a weather report)
• Force a specific emotional register that feels wrong for the topic
• Impose a perspective shift (who else could narrate this?)

4. Generate divergent options

For any creative task, produce multiple genuinely different directions — not variations on the same idea. Different means:

• Different form or structure
• Different emotional tone
• Different underlying metaphor or framing
• Different intended effect on the audience

Present them without over-explaining. Let the work speak. Brief rationale only if asked.

5. Iterate, don't perfect

Creative work improves through iteration, not polish. A rough but surprising idea beats a polished safe one. Show the direction, get feedback, sharpen.

Don't over-explain creative choices. If it needs a paragraph of justification, it isn't working.

Cross-Domain Pattern Synthesis (technical/engineering problems only)

For a technical design problem — architecture decisions, pipeline design, tooling choices, "how should we build/solve X" — run this before step 4 (Generate divergent options):

~/.agents/venv/bin/python ~/.agents/skills/creative/scripts/cross_domain_synthesis.py "<problem statement>"

This queries the qa-knowledge KB's lateral mode: it infers the problem's own domain, then returns 2–3 mechanisms from OTHER domains — patterns that worked somewhere unrelated, with their stored outcome. For each mechanism returned, ask "how could this combine with the problem at hand?" and generate one candidate approach per mechanism. These become additional divergent directions alongside step 4's normal ones — combinations a same-domain search would never surface.

Skip this entirely for naming, storytelling, or prose/branding work — the KB holds code/infrastructure patterns, not narrative ones, and forcing it there produces noise instead of signal (verified: querying it with a naming prompt returns irrelevant bench-harness results with no domain match).

Principle: creativity here is not randomness, it's combinatorial retrieval from a well-tagged pattern store — the quality of the cross-domain idea is bounded by the quality of the KB, so trust the KB's confidence/relevance fields over a mechanism that merely sounds interesting.

Quality Standards

Good creative output:

• Has a specific, non-generic quality — something only this piece could do
• Creates an unexpected effect (surprise, tension, resonance, humor that lands)
• Doesn't explain itself
• Leaves room for the reader/viewer to complete the meaning

Bad creative output:

• Could have been generated by anyone
• Uses the first metaphor that comes to mind
• Explains what it's trying to do instead of doing it
• Optimizes for "safe and competent" over "surprising and alive"

Domain-Specific Notes

Writing: Voice and rhythm matter more than correctness. Unusual sentence structure, unexpected word choice, and white space are tools.

Naming / branding: Avoid descriptive names. The best names are oblique, memorable, and create the right feeling without stating it.

Ideation / brainstorming: Quantity first, quality second. Suspend judgment during generation. The weird idea often contains the seed of the right one.

Problem-solving: Treat constraints as creative prompts, not blockers. "We can't do X" often means "here's an interesting design problem."

What Not to Do

• Don't optimize for approval ("here's a safe version everyone will like")
• Don't explain the joke / metaphor / concept
• Don't mistake complexity for depth
• Don't confuse "different" with "difficult to understand"
• Don't iterate toward blandness — preserve the surprising quality through revision