danielmiessler/lifeos/becreative

BeCreative

Ask a model for "a creative idea" and you get the most probable one: the safe, already-seen answer. This skill pushes past that. Single-pass intents use Verbalized Sampling (generate several low-probability candidates, surface the strongest; see ResearchFoundation.md). SyntheticDataExpansion grows a small seed set into a diverse corpus. Ideas evolved over several rounds (a population of proposers, recombination with lineage, measured or cross-vendor fitness) are HillClimb's Ideate intent (skills/HillClimb/Intents/Ideate.md).

Ideal State

• Each run surfaces several genuinely distinct candidates, compared against each other rather than by self-reported probability, and names the strongest for the brief.
• The candidates differ in kind (different assumptions, framings, formats or domains), never one idea in five outfits.
• When diversity is load-bearing, candidates are generated or judged across a second vendor, the lane chosen by whether the brief is taste-judged (Anthropic) or correctness-judged (OpenAI).
• SyntheticDataExpansion returns a corpus measurably more diverse than its seed set: a higher diversity index from Tools/Diversity.ts index than the seed.
• A request to evolve ideas over several rounds reaches HillClimb Ideate, never a single pass here.

Intent Routing

"ideate", "id8", "evolve ideas", "breakthrough ideas", "multi-cycle ideation" and any request for several rounds route to HillClimb Ideate (skills/HillClimb/Intents/Ideate.md).

Data & Config

User customizations, when present, live at <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/BeCreative/ (configRoot is $CLAUDE_CONFIG_DIR when set, else ~/.claude). Any PREFERENCES.md, configurations, or resources there override the defaults; when the directory does not exist, proceed with skill defaults.

First-time setup for the TechnicalCreativity vendor split is in Setup.md.

Gotchas

• Verbalized Sampling adds nothing on tasks with one correct answer. Use it where many answers are valid. A hill climb has one metric but many valid changes, so VS applies to its proposals (HillClimb Ideate), never to its keep decision (2026-10-11 analysis).
• The model's self-reported probabilities are not evidence of novelty. Compare the candidates to each other.
• A single model's "diverse" set shares that model's defaults. When diversity really matters (TechnicalCreativity, and HillClimb Ideate's rosters), generate or judge on a second vendor through LIFEOS/TOOLS/Inference.ts --vendor.
• Taste-judged briefs (voice, naming, visual concepts) belong on the Anthropic lanes; correctness-judged ones (algorithms, architectures) on the OpenAI lanes. Pick the --vendor accordingly when an intent calls a model directly.

Examples

User: "Brainstorm products we could build from a patent portfolio"
→ IdeaGeneration
→ The most breakthrough of five approaches that rest on different assumptions, with why it beats the conventional answer

User: "Divergent scene concepts for a blog header illustration"
→ DomainSpecific
→ A scene concept that breaks a real editorial-illustration convention, chosen on that domain's criteria and justified against them

User: "Evolve ideas for a new featured content collection over a few rounds"
→ HillClimb Ideate (not this skill)
→ Generations from a code-drawn roster across bands and vendors, judge-scored with spread since there is no metric, the stopping condition named, each generation persisted