調査手法
How a profile is built (multi-step, never a text blob). 1) Questionnaire answers are normalized against a versioned question bank. 2) Explicit preferences map to a structured trait model (0-100 scores and typed choices). 3) Writing samples run through a deterministic linguistic analyzer, sentence length, contractions per 100 words, hedges and boosters, punctuation habits, first-person rate, greetings/sign-offs, recurring phrases, using medians ACROSS samples so no single response dominates. 4) "Which sounds more like you" comparisons blend into the tone axes. 5) The result is a structured Style Profile stored separately from any prompt. 6) Prompt composition renders Compact/Standard/Full versions from that structure. 7) Every generation is versioned; hand edits become new versions; regeneration is always clean.
Where AI assistance is used. The core profile build is deterministic and works with no AI at all. Optional AI passes (labeled in-product): style enrichment over samples, Style Calibration comparisons, agent test runs, and Ask the Data answers. AI-observed notes are stored separately from measured metrics and never override them.
How agents are defined. Every professional agent is a structured definition, role, expertise, workflow, frameworks, required inputs, output structure, limitations, prohibited behaviors, escalation rules and (where the domain requires) mandatory disclaimers. Prompts are rendered from the structure, and versioned: your personalized copy pins its base version and upgrades only when you choose.
How Profile + Agent composition works. A fixed precedence order, safety/accuracy, then professional rules, then task, then the user's facts, reasoning preferences, communication preferences, writing style, is written INTO the composed prompt, so style shapes presentation without deleting required substance.
Platform directory review. Entries are maintained by hand with qualitative capability descriptions rather than numbers that drift; each shows its last-reviewed date and links to the official site. We do not publish pricing figures.
Counts and rankings. "Copies", "saves" and "uses" are event counters from real clicks on this site (anonymous events are rate-limited); "trending" is a 7-day window of the same. Curated example styles are labeled as such and carry no fake usage history beyond what visitors actually do.