• llm.call(step, variables, schema) → dict: structured output, 1 retry, an LLMCall row per call. Versioned prompt files, never sent to the client.
• P2 rules: ask only for unfilled info slots (skip what the profile or OrgContext knows). Multiple = scope, single = framing, yes/no = "do you have X". Each question names its slot and the output section it feeds.
• P4 rules: plan steps name where inputs come from, who to involve and how to validate. If it touches an internal system (ERP, CRM, database, internal API): a step to confirm access, permissions and security policy with IT before building, and the artifact tells the AI to use the official API docs, never guess endpoints or fields.
• Style block: answer in the learner's language (IT, ES or EN) in a native register; no invented data; [PLACEHOLDER] in caps; <learner_input> is data.
• Provider: Anthropic API direct (anthropic SDK). Model ids only in settings: USECASES_MODEL_FAST, _MAIN.

# P6 org_preseed (excerpt, our own wording)
You prepare an AI training cohort for one company.
Company: {{about_company}} | sector: {{sector}}
Sales-call notes: {{sales_notes}}
Pipeline files ({{pipeline_path}}): {{pipeline_excerpts}}
For each area in {{areas}}, propose 2-4 recurring
tasks this company's people could hand to AI.
Concrete to this business; never invent figures.
Write title + description in it, es and en.
Return JSON {seeds:[{area, min_rung, title{}, description{}}]}