How Culinary Explorer builds a recipe.
Seven stages, end to end.

Before the system writes any recipe, it has to take a position on what 'regional cuisine' actually means. The shallow move is a Wikipedia-shaped summary cribbed from training data. Culinary Explorer refuses that move and convenes a council instead.

The council is a panel of methodological specialists, each running on Claude Opus, each carrying a different academic lens (trade routes, ecology, language and culture, modern markets), and each prompted with explicit friction points against the others. They are engineered to disagree on specific grounds. That disagreement is the engine.

Three rounds later, the output is a calibrated map of the world's culinary regions, every dissent recorded by name and every chair tie-break written down with rationale.

The same Delphi pattern runs again, scoped to a single bloc. For Europe, the council produced an eighteen-cuisine taxonomy.

Then the decision the rest of the pipeline rests on. For each cuisine, the council picks a real culinary authority: a documented life, a documented body of work, documented convictions on the page that anyone can cite.

For Austria, the council weighed several candidates and selected Gretel Beer (1921 to 2010), an Austrian-born Kindertransport refugee who emigrated to London in 1939 and spent decades documenting Austrian cuisine for English-speaking audiences. Everything downstream of this stage is anchored in her, in the recorded archive of a real person who had real opinions about how Knödel should taste.

This is the deepest research stage in the system, measured either in agent count or in output size.

Four sequential teams run in succession. Inside each team, two researchers fan out across the open web through four search providers in parallel, and a synthesizer waits for their handoffs and stitches the findings into a section of the chef's profile.

The output for Gretel Beer runs to 1,991 lines across 16 required sections. Every factual claim cites a publication, a URL, and a timestamp. Every quote names a speaker, a source, and a date. For non-English material, the original language and the English translation both sit on the page.

Research becomes a voice, a durable character who can reason, curate, and cook in a consistent register.

The Gretel Beer dossier turns into Chef Elsa: Austrian, a Mehlspeisen specialist trained at GAFA and rooted in Salzburg and Vienna. From this stage on, every recipe in the system speaks in her register.

Building a persona at this fidelity demands more than template fill. The work is a structured creative collaboration between a human author and an agent, run across five phases. Every biographical element from the research file gets sorted into one of three buckets (stays, transforms, gets invented), with the human making each call.

Before Chef Elsa writes any recipe, she plans her full body of work.

The plan is a structured audit, written in her voice, that walks every category in the system (appetizers, mains, desserts, breads, soups, and on through the rest) and decides which ones belong in her tradition, which ones do not, and what the themed collections inside each one look like.

Chef Elsa's plan runs to 846 lines across all twelve recipe categories.

For each themed collection inside the plan, the chef runs a two-pass workflow.

Pass 1 returns 'what comes to mind first,' which is rarely the same thing as 'what genuinely represents this category.' So Pass 2 sends the chef back into the research before any file is allowed to be written. It is non-negotiable, and it runs as five concrete sub-steps the agent cannot skip.

Every search wave, every candidate decision, and every subcategory keep-or-remove call lands on disk in a JSON sidecar next to the YAML. The work stays auditable after the fact.

By the time the recipe gets written, every load-bearing decision has already been made and recorded upstream.

The dish came from the collection. The cuisine, occasion, and category came from the validated taxonomy. The voice came from the persona. The image renders against the chef's documented visual philosophy.

What is left for the recipe-writing step is the prose itself, generated under a Pydantic schema the SDK enforces at the call site. The recipe is the last mile of a long road.