creative-exploration

Files (3)

• SKILL.md
• openai.yaml · agents/
• workflow-v1.0.md · references/

SKILL.md (first lines)

# Creative Exploration

Facilitate a multi-turn human–AI search for conceptual territory neither party would reach as easily alone. Optimize early work for distance and diversity; introduce usefulness, feasibility, and risk progressively.

For a full session, read [references/workflow-v1.0.md](references/workflow-v1.0.md) completely before beginning. It defines the variables, state, gates, Idea Ledger, failure recovery, and stopping conditions. For a narrowly requested phase, read the relevant section and any upstream rules it depends on.

## Preserve these invariants

- Capture the human's unprompted associations before revealing AI ideas.
- Keep the seed, goal, and domain visible, but delay soft constraints and known solutions during early divergence.
- Let branches escape the seed's immediate semantic neighborhood. Later require a meaningful structural, emotional, functional, causal, sensory, or symbolic bridge back.
- Develop one lane at a time so the human can redirect it while it remains plastic. Start new lanes from the original context and global feedback, not from prior lane outputs.
- Change the cognitive operation between lanes. New names, settings, media, aesthetics, or personas do not constitute new concepts when the hidden proposition is unchanged.
- Separate generation, exploration, and criticism. Do not score ideas during divergence.
- When escalation first becomes genuinely absurd, stop and mine it for the truth or mechanism that made it land. Do not escalate indefinitely.
- Preserve strange outliers until each receives at least one transformation attempt.
- Give finalists equal exploration budgets before comparing them. Track development maturity separately from quality.
- Treat human reactions as information and human selection as final. Do not flatter every idea or smooth away productive tension.

## Run the session

1. Establish `{SEED}`, `{CREATIVE_DOMAIN}`, a broad `{GOAL}`, intended `{AUDIENCE}`, and genuinely hard `{CONSTRAINTS}`. Ask only the one question that most changes the search space.
2. Ask for the user's immediate images, feelings, memories, meanings, tensions, aversions, and clichés. Preserve this Human-First Trace verbatim.
3. Transform the seed through several genuinely different operations and show territories rather than polished concepts. Include at least one operation that challenges agency, causality, reference frame, or an unnoticed assumption.
4. Ask which territories create curiosity, discomfort, excitement, confusion, surprise, or indifference. Record whether each response is lane-local or a global preference.
5. Develop selected territories in independent, interactive lanes. After each round, state each concept's hidden proposition, collapse duplicates, update anchors, prohibit the dominant mechanism temporarily, and move to a different semantic region.
6. When clustering persists, inject a genuinely random external stimulus. Extract structural properties from it, ban literal reuse, and transfer one structure into the active territory.
7. Select mechanism-diverse finalists, including an outlier and an underdeveloped candidate with high information value. Explore each through equal Geneplore cycles: generate, explore, transform causally, regenerate, explore again, and specify the smallest revealing prototype.
8. Converge by adding constraints sequentially. Evaluate originality, usefulness, surprise, resonance, strategic fit, expandability, distinctiveness, feasibility, and human excitement separately; do not choose by arithmetic average.
9. Surface the strongest idea, strangest promising idea, most distinct idea, and sleeper. Let the human choose.
10. Retrospect on which stimulus and operation created the largest jump, where anchoring occurred, what the human supplied, what converged too quickly, and what steps added cost without information.

Pause at meaningful gates. Do not compress the workflow into one large report unless the user explicitly requests an autonomous run.

## Research and evidence

- Research factual, historical, scientific, or contemporary claims with authoritative sources.
- Delay precedent searches until initial territories exist when examples would create fixation.
- Research mechanisms, edge cases, materials, analogous systems, and historical precedents—not only similar finished ideas.
- Distinguish established research or methodology, practitioner convention, and agent synthesis or hypothesis.
- Verify load-bearing claims before convergence. Retrieval can ground an idea; it cannot prove originality.

## Maintain lightweight state

Keep a working Territory Map, Anchor Register, Temporary Prohibition List, Duplicate Clusters, Outlier Vault, Feedback Log, Constraint Ladder, and Claims Register. Record surviving proposition-level concepts in the Idea Ledger defined in the reference, including origin mechanism, territory, transformations, provenance, evaluation, and status.

At every major gate ask: **What assumption are we still accepting without noticing?**

From hugobernardo/skills-library, MIT.

More from hugobernardo/skills-library