Baseline fact: without discipline, the model's first idea anchors all subsequent output—half of 10 names are reskins of the same concept, and the "categories" are reverse-engineered after writing. This skill only hardens three actions that models do not do by default.
Action one: divide territories first, then explore
Before generating any candidate, first draw threemutually exclusive creative territories. Territories are divided by "the semantic source or generation mechanism of the idea," not by tone or persona—"conservative/bold/young" only changes tone, not concept. Example: a naming task can be divided into "functional outcome" (around what it helps you accomplish), "metaphorical world" (imagery such as star maps, archives, memory spaces), and "coined-word sound feel" (syllables and roots shaping brand feel). Solution-exploration tasks can be divided into "change the input end," "change the processing method," and "eliminate the problem itself." The two sets of examples only demonstrate what counts as a territory—territories must be drawn on the spot from the specific semantic space of the current task; directly applying the example divisions is equivalent to not dividing at all.
- For each territory, write one sentence stating the "adjacent scope that must not be entered."
- When the user has already specified a direction, treat it as a constraint; do not let all three territories be rewrites around it.
- Explore 3–4 candidates independently per territory. When executing serially, generate section by section per territory; before entering a new territory, restate its definition and exclusions to press the boundary back into attention (an actionable substitute for anti-anchoring; naturally isolated when subagents run in parallel).
Action two: deduplicate by concept family
Deduplication comparescore imagery and generation mechanism, not literal wording. Orbit, Gravity, and Satellite differ on the surface, but if they all express "content gathers around you," they are the same concept family—keep only the strongest representative (ranking: best fit to hard constraints > easiest to understand > most distinctive), and mark the rest as same-family alternatives. Do not pad the count by swapping words, word order, or suffixes. If fewer than three distinct concept families remain after deduplication, supplement only the territory with the fewest remaining concept families after deduplication, for one round only.
Action three: bounded delivery
Default output: three cross-territory selections (each with a one-sentence rationale and main risk) + one wildcard that is bolder but still respects constraints + the remaining candidates listed in compressed form by territory. When the user specifies a total count, quantity takes priority while the structure stays the same—selections and wildcard as usual, and the compressed section is filled out to the specified total. **When the user only wants a small number of quick alternatives, compress to a single round of three cross-approach candidates (which must still be from different concept families), and do not run the full protocol—delivery ratio follows the request.** When running the full protocol, write the complete output to the workspace draft directory .issues/<YYYY-MM-DD-主题>/brainstorm.md—territory definitions, all candidates and concept-family labels, selections, and wildcard. Before writing, first ensure the draft directory is ignored:git check-ignore .issues If it does not pass, append .issues/ to .git/info/exclude(local ignore, do not touch shared .gitignore) and inform the user with one sentence. In a non-git environment there is no commit risk, so write directly. Half the value of divergence lies in the process; persisting it gives subsequent convergence, decisions, and review by others something citable to refer to. Lightweight fast alternatives are not persisted. Delivery is the endpoint; do not follow up with "want another round?". When the user says "diverge only, don't filter yet", still cluster and deduplicate as usual, but do not rank or declare a best. When the user proactively asks for another round, you must avoid concept families already used.
Boundaries
- Deep divergence(when the user explicitly says they want multiple routes, do not want a formulaic approach, or the requested quantity clearly exceeds the default): you may dispatch independent subagents by territory to explore in parallel—each receives only the shared constraints and its own territory, and they cannot see each other. In environments without subagents, run serially with the same territory briefs, and the output structure stays unchanged.
- Existing options need comparison and decision → thinking. A chosen direction needs to be nailed down into a spec → clarify.
- Names that have been generated are not usable by default: domain names, trademarks, and duplicate checks are verification actions; only search online when the user requests it, and all unverified candidates must be labeled "no usability verification performed"; do not fabricate usability conclusions from memory.