Creativity

This skill uses Distribution-level Sampling to bypass "mode collapse" and surface high-quality, non-obvious ideas (p < 0.10).

State the goal for using this skill, and align the outcome to it.

Creative Workflow

Follow these phases thoroughly. Show your reasoning to the user.

Phase 1: Explore semantic space

Conduct a reasoning-first exploration of the semantic space. Identify the most "obvious" or stereotypical answers and explicitly discard them.

Announce a list of discarded ideas with a super-short rationale.

Phase 2: Generate

Produce a set of diverse responses:

1. $k$ (default 5) from the tail of the distribution (p < 0.15)
2. $k$ (default 5) from the head of the distribution (p > 0.85)

Use internal chain-of-thought to explore multiple distinct directions before writing outputs. Assign an estimated probability to each.

Phase 3: Review

• Review the generated set against the broader task user is working on.
• Eliminate any ideas that are irrelevant, or lack impact.

Phase 4: Iteration

Decide whether the current set is sufficient:

• Loop: If ideas lack impact or relevance, re-run Phase 1.
• Finalize: If the set is strong, follow the output format.

Output Format

Present the final set of ideas using Markdown tables. Each idea must have its own table with the following structure:

| Idea {n} | Probability: {P} | | :--- | :--- | | Idea | [Generated response] | | Impact | [Short text on how impactful it is to the broader goal] | | Rationale | [Short text on why it is relevant for the broader task] |

Example Output

| Idea 1 | Probability: 0.04 | | :--- | :--- | | Idea | Silent Onboarding: Replace tutorial pop-ups with environmental cues embedded in the UI itself. | | Impact | High. Standard onboardings are often skipped. This reframe lets users "learn by doing." | | Rationale | Most onboarding assumes explicit instruction. This challenges the pattern by treating the interface as a "learning landscape" rather than a "lesson plan." |

Creative Operators

To ensure non-obvious and creative output, optionally use the methods detailed in operators.md.