Innovation Assistant Skill

A pure-Markdown skill that turns an AI agent into a structured innovation engine. It encodes two research-backed methodologies — SIT (Systematic Inventive Thinking, from the template research of Goldenberg, Mazursky & Solomon) and Think Bigger (Sheena Iyengar's Choice Mapping) — as executable pipelines with hard validators, plus a judgment layer of ten evidence-backed axioms.

The idea

Innovation is widely believed to be a stroke of genius. The research record says otherwise: most successful innovations are new, useful combinations of parts that already existed, and the combining can be done by procedure — fixed transformation templates, cross-domain precedent search, mandatory randomized recombination, saturation-based stopping rules. Procedures are exactly what AI agents execute well. What an off-the-shelf model lacks is the procedure itself: left alone, it produces eloquent trend summaries. This skill supplies the procedure.

The division of labor is deliberate. The agent does the legwork: decomposition, precedent search, mechanical transformation, combination, scoring, audit trails. The human keeps what cannot be delegated: choosing the problem, stating their wants, taste calls on the shortlist, and real-world validation.

What's in the box

Path	What it is
skills/innovation_assistant.md	Root skill: routes a task to the right pipeline (or honestly to neither), hard rules, human-in-the-loop nodes, acceptance criteria, output spec
skills/sit_pipeline.md	SIT: product decomposition schema, six elementary operators, five templates with fixed operator sequences, dependency matrix, function-follows-form scan
skills/think_bigger_pipeline.md	Think Bigger: six steps from problem choice to third-eye rehearsal, with numeric validators (≤5 subproblems, 5 tactics/row with ≥3 out-of-domain, seeded random combination rounds...)
axioms/axioms.md	Ten axioms — the judgment layer that explains why the rules exist, each backed by primary literature
experiments/ai_interface_innovation/	A/B evidence: the same task run by the same model with and without the skill, plus a third-party evaluation

Does it actually change anything?

We ran a paired experiment: two Claude Opus agents, identical task ("what interface innovations are still possible for AI agents?"), identical input article and tool access; one loaded the skill, one didn't.

The baseline agent produced an excellent market survey whose recommendations were mostly "join existing trends". The skill agent produced buildable candidates with derivation chains — its winning idea combines eBay-style reputation scoring, SAE-style autonomy levels, and bank-style risk thresholds into a "trust ladder" interface, a mechanism assembled from 15 out-of-domain precedents the baseline never touched. Both agents independently confirmed the market gap the winner sits in, which is the kind of convergence you want. Full reports and the evaluation are in experiments/ai_interface_innovation/ (reports in Chinese; the setup and evaluation verdict are summarized in the experiment README).

One pair is not statistics. It is, however, fully inspectable: every idea in the skill report carries its lineage, so you can audit exactly where the structure helped.

Installing the skill for your agent

The skill is loose Markdown — no packaging, no runtime. Hand this repository's URL to Claude Code, Codex, Cursor, OpenCode, or any coding agent and ask it to install the skill into your workspace:

1. Clone or vendor this repo somewhere your agent can read.
2. Add a pointer to skills/innovation_assistant.md in your workspace's skill discovery chain — AGENTS.md, CLAUDE.md, or a skill index like rules/skills/INDEX.md if you have one. Expose only the root skill; it links to the rest.
3. Invoke it by asking your agent to "run structured innovation on ..." — the root skill routes from there.

The pipelines assume the agent has web search available for precedent hunting (any search tool works); everything else is plain reasoning.

Honest limitations

• The famous numbers behind these methods (≈70% of successful products match a template; 89–92% classification accuracy) are retrospective statistics with sampling bias, not forward success rates. The skill forbids presenting them otherwise.
• The methods excel at combinational and improvement innovation on definable problems. For paradigm-level questions the router's honest answer is "not this tool".
• Real-world validation (showing ideas to actual humans) cannot be simulated by the agent. The skill prepares the feedback kit; you still have to talk to people.

Sources

Goldenberg, Mazursky & Solomon (1999, J. Marketing Research); Goldenberg, Lehmann & Mazursky (2001, Management Science); Sagiv, Arieli, Goldenberg & Goldschmidt (2010, J. Organizational Behavior); Diehl & Stroebe (1987, JPSP); Iyengar, Think Bigger (Columbia University Press, 2023); Schumpeter (1912); Weitzman (1998, QJE). Primary papers were read in full; the book via a multi-source verified digest.

License

MIT