CrossGen — Skill de Geração de Soluções por Mapeamento Estrutural
mohamedAtoui/crossgen-skill
Mapeia mecanismos entre domínios para resolver problemas difíceis com inovação estruturada
Ver no GitHubUma skill Claude que gera soluções genuinamente inovadoras mapeando estruturas e mecanismos entre domínios distantes (biologia, física, ecologia, artes). Executa um pipeline de 9 estágios fundamentado em Teoria de Mapeamento Estrutural de Gentner, TRIZ e biomimética, produzindo soluções testáveis com previsões falsificáveis.
Instalação
npx convez-ai mohamedAtoui/crossgen-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerar soluções inovadoras para problemas técnicos e estratégicos complexos
- Aplicar biomimética e princípios universais (TRIZ) na resolução de contradições de engenharia
- Desenvolver ideias com testabilidade garantida e mapeamento causal profundo
Veredito

Skill rigorosa e bem-arquitetada que vai além do brainstorming superficial, com ênfase em mapeamento causal e previsões falsificáveis. Exige investimento de tempo para implementação, mas entrega valor real em problemas de alta complexidade.
README
A tradução para o português deste README ainda está na fila.
What it does
You give Claude a hard problem. CrossGen runs a 9-stage analogical-reasoning pipeline that:
- Reformulates the problem 5 different ways (inversion, constraint-relaxation, abstraction, provocation, first-principles) and explores 3 in parallel.
- Decomposes each reframing into functions, constraints, contradictions, causal chains, and Meadows-level leverage points.
- Abstracts through 4 simultaneous lenses: SAPPhIRE (scientific principles), Biologize (biomimicry), WordTree (semantic expansion), TRIZ (40 inventive principles + contradiction matrix).
- Expands to 6–8 high-distance candidate domains across biology, physics, ecology, arts, earth sciences.
- Mines each domain for a specific mechanism, then maps it to the target using Gentner's Structure-Mapping Theory — relations, not surface features. Five+ one-to-one mappings, connected causal chains, mandatory "where it breaks" with ≥3 weak spots.
- Synthesizes each analogy into a concrete solution with numbered implementation steps, candidate inferences, testable falsifiable predictions, and research pointers.
- Verifies each solution against existing work; pushes weak ones toward genuine novelty.
- Evaluates on six calibrated dimensions (novelty, verified-novelty, surprise, feasibility, structural depth, actionability) with a same-domain penalty.
- Evolves the top ideas through mutation (mechanism-swap, scale-shift, constraint-flip, domain-blend), recombination, and critique-and-strengthen.
The output is a ranked list of solutions with mechanism, scores, numbered concrete steps, testable predictions, honest failure modes, and academic-search pointers.
What makes it different from "ask Claude to brainstorm"
- Map relations, not surface features. "Both have networks" is banned. Every mapping must trace HOW the source mechanism operates step-by-step against the target's causal chain.
- "Where it breaks" is mandatory. Every analogy must list ≥3 specific weak spots. Every concrete step must explicitly address one. Inspiration-porn analogies don't ship.
- Testable predictions are non-negotiable. Every solution outputs ≥2 falsifiable predictions with quantitative thresholds — the most valuable artifact of the entire pipeline.
- TRIZ is deterministic. The 40 principles and contradiction matrix are looked up, not hallucinated.
- 40 universal principles × 42 curated domains are bundled as reference files Claude loads on demand — feedback loops, phase transitions, self-organization, catalysis, mycelial networks, stigmergy, allosteric regulation, capillary action, metamorphosis, quenching, trophic cascades, chirality, dormancy…
- Honest scoring. Same-domain penalty (×0.5). Calibrated rubric. Distribution sanity-check. Most scores belong in the 0.4–0.7 band.
Install on skills.sh
- Upload this folder as a Skill to skills.sh — the entire repo (SKILL.md + references + examples + assets) is the skill package.
- Once installed, Claude will trigger it automatically when you ask things like:
- "How can we solve X?" (where X is a hard problem with trade-offs)
- "Give me a creative approach to Y"
- "Use biomimicry / TRIZ / lateral thinking on Z"
- "Find non-obvious solutions to W"
- You can also invoke explicitly: /crossgen "".
Or test it locally:
git clone https://github.com/<your-user>/crossgen-skill
cd crossgen-skill
# In Claude Code, point to this directory as a local skill
How it works (file layout)
crossgen-skill/
├── SKILL.md ← entry point Claude reads (frontmatter + workflow)
├── references/
│ ├── structure-mapping.md Gentner's theory operationalized — banned vs required patterns
│ ├── prompts.md Verbatim stage prompts (each of the 9 stages)
│ ├── universal-principles.md 40 cross-domain principles, tagged with domains
│ ├── domain-catalog.md 42 candidate domains across 8 categories
│ ├── triz.md 40 TRIZ inventive principles + contradiction matrix
│ ├── scoring-rubric.md Calibrated scoring with examples
│ └── output-template.md Final report format
├── examples/
│ └── extreme-heat-cities.md Abridged real run — top 5 solutions, quality bar
├── assets/
│ └── banner.png
├── LICENSE MIT
└── README.md this file
Reference files are loaded progressively — Claude only reads them when relevant (e.g., structure-mapping.md before Stage 4 if rusty on Gentner; triz.md during Stage 2d).
Example
Problem: "How can cities remain functional when extreme heat events exceed 50°C for weeks?"
Top solution (score 0.77, distilled): Termite mound architecture + Nitinol shape-memory-alloy passive actuators. Replace electric AC micro-pumps with SMA wire loops that mechanically dilate channels above 35°C. The system is anti-fragile — the worse the heat event, the stronger the actuating force — and the natural ±0.5°C tolerance in production Nitinol provides ±18-minute desynchronization across a 500-building precinct without any firmware, compliance, or governance layer.
Testable prediction: Thermosiphon circulation velocity in a 1.8m vertical loop scales as v ∝ (ΔT)^0.5 with coefficient ~0.04 m/s·K^−0.5 — predicting v ≈ 0.19 m/s at ΔT = 20°C, sufficient to maintain ≥60 W/m² discharge flux pumpless.
Full example: examples/extreme-heat-cities.md. The original Python pipeline's full 40-solution output is in the source repo.
When NOT to use this skill
- Factual lookups
- Code generation
- Summarization
- Questions where the "creative" framing is itself the wrong move (regulatory compliance, security incident response, hard-deadline patch fixes)
If you want one off-the-cuff novel idea, regular Claude works fine. CrossGen earns its compute on problems where trade-offs are sharp, conventional answers are exhausted, and you need rigor — not just inspiration.
Lineage
This skill is the distilled methodology from the CrossGen Python pipeline (~3K lines of Python orchestrating Claude across 9 stages with quality gates, parallel paths, deterministic TRIZ lookup, and automatic novelty verification). The Python version produces machine-readable JSON; this Skill version produces equivalent reasoning natively inside Claude — no API plumbing required.
The intellectual grounding:
- Gentner, D. (1983) — Structure-Mapping: A Theoretical Framework for Analogy. Cognitive Science 7(2).
- Altshuller, G. (1984) — Creativity as an Exact Science. TRIZ foundational text.
- de Bono, E. (1970) — Lateral Thinking.
- Meadows, D. (1999) — Leverage Points: Places to Intervene in a System.
- Benyus, J. (1997) — Biomimicry: Innovation Inspired by Nature.
- Hofstadter, D. & Sander, E. (2013) — Surfaces and Essences.
License
MIT