# How a CulinaryExplorer Recipe Gets Made > A canonical description of the research and curation machinery behind every recipe we publish. --- ## Purpose This document describes the pipeline by which a recipe arrives on CulinaryExplorer. It is a single-source reference for how the system works: the stages, the artifacts produced at each stage, and the engineering that drives them. --- ## The pipeline in one picture ``` Global Cuisine Council (defines world culinary regions) │ ▼ Regional Cuisine Council (defines countries / cuisines in a region) │ ▼ Chef Research Profile (deep research on a real culinary authority) │ ▼ Chef Persona (a fictional chef, grounded in that research) │ ▼ Collection Plan (the chef's full recipe coverage strategy) │ ▼ Recipe Collection (the curated dish list for one theme) │ ▼ Published Recipe ``` Every stage is a research artifact in its own right, with its own audit trail. By the time we write a recipe, the dish has already been justified at six levels above it. --- ## Stage 1 — The Global Cuisine Council **The question:** What are the world's culinary regions, and in what order should we cover them? **The machinery:** A five-member panel of AI expert agents, each one a fully-instantiated persona with an academic background, a doctoral thesis, a list of published works, and a defined methodological lens. The lenses are deliberately incommensurable: - A trade-route and diaspora historian who insists "regional cuisines" are residue of historical migration - An agricultural ecologist who maps cuisine onto rice belts, wheat belts, and biomes - A linguistic-cultural anthropologist who ties cuisine to language families and religious dietary law - A modern-markets analyst who weighs cookbook sales, restaurant counts, and English-language search volume - A neutral chair — a Pacific-based UNESCO heritage methodologist with no home team in the four lenses The personas are engineered for **friction**, not consensus. Each specialist has documented disagreements with at least one other specialist baked into their persona. This is the design. The panel runs three rounds under multi-agent orchestration with structured message-passing between agents: 1. **Round 1 — Parallel research.** Each specialist independently maps the entire world through their own lens, using a multi-provider search tool that fans out to several search engines simultaneously and cross-references results. Every claim must be cited; training-data answers are not accepted. Each specialist returns a bloc list, member-country lists, and an explicit list of borders they consider "soft." 2. **Round 2 — Cross-examination.** The chair compiles a conflict matrix showing, country by country, where the four maps disagree. The panel then debates each contested boundary and a list of mandatory edge cases (Mediterranean: bloc or overlay? Where does Eastern Europe begin? Is Mexico Latin American or North American?). Specialists argue in 200-300 word turns; each of the other three responds. 3. **Round 3 — Calibration.** The chair drives convergence. Consensus rules: 3-of-4 specialists agreeing passes; 2-2 splits go to the chair with written justification recorded. True methodological deadlocks (where two valid lenses produce incompatible answers) are routed to an `overlays:` escape hatch in the schema rather than forced into a false consensus. Every round is persisted to disk as a separate working artifact. Dissents are recorded by name. The output is a priority-ranked map of regional blocs and a one-page council brief for each — the seed document for the next council down. --- ## Stage 2 — Regional Cuisine Council The same Delphi-style panel pattern runs again at the regional level, with new specialists scoped to one bloc. The European Cuisine Council, for example, identified the eighteen cuisines of Western and Northern Europe and — critically — selected a real **culinary authority** for each one. For Austria, the council weighed several candidates and selected **Gretel Beer**: an Austrian-born journalist and food writer who emigrated to London during the war and spent a career documenting the Viennese kitchen for an English-speaking audience. The selection rationale, the alternatives considered, and the dissents are all recorded. Selecting a real authority is the load-bearing decision of this stage. It anchors everything downstream in a documented body of work, a documented set of convictions, and a documented life — not a generic regional summary. --- ## Stage 3 — Chef Research Profile This is the deepest research stage in the system. The output is a sixteen-section research document on the selected authority. The machinery that produces it is what makes "deep research" mean something. **Four sequential agent teams.** Each team contains two researchers and a synthesizer, all spawned in parallel. Teams cover different section clusters: 1. **Foundation** — biography, philosophy, legacy 2. **Voice & character** — personality, anecdotes, writing style, fifteen to twenty-five attributed quotes 3. **Visual & cultural** — anti-identity, recipe content samples, narrative voice, photographic style, cultural material culture (ceramics, vessels, surfaces, color signatures) 4. **Final synthesis** — signature dishes, persona synthesis notes, source bibliography **Concurrent multi-provider search.** Researchers use a search tool that calls several engines in parallel and returns cross-referenced results in one pass. Each researcher runs three to four search waves per section. **Persistence-first delivery protocol.** Findings are written to disk **before** any agent-to-agent handoff. Researchers write their full findings to a `.research-findings/` audit directory, then send a short pointer message to the synthesizer. If the message bus fails for any reason, the synthesizer can read the file directly. This is fault-tolerant by design. **Watchdog and receipt verification.** The orchestrator actively monitors researcher progress; stalled researchers get pinged and, if non-responsive, replaced. After researchers report complete, the orchestrator asks the synthesizer to confirm the message count — missing handoffs are forwarded by hand. No silent failures. **Citation and attribution requirements (non-negotiable):** - Every factual claim cites a source - Every quote has speaker, source, and date attribution - For non-English subjects, both original-language and English translations are captured - Unverified claims are explicitly flagged `[UNVERIFIED]` **Synthesis discipline.** The synthesizer reads the existing research document at the start of its turn (so it doesn't overwrite prior teams' work), integrates the new findings, and writes its assigned sections. The Team 4 synthesizer additionally reads the entire completed document and produces a persona recommendation block — recommended archetype with rationale, voice compass, key pitfalls, differentiation notes from existing personas, and open questions. The result for Gretel Beer is a research document with hundreds of citations, dozens of attributed quotes, a documented analysis of her writing style and tonal range, and a complete picture of Austrian cultural material culture — ceramics, kitchen surfaces, color signatures, plating traditions. This document is the factual foundation for the persona. --- ## Stage 4 — The Chef Persona The persona is where research becomes voice. This is **not** a template fill. It is a structured creative collaboration between a human and an AI agent that runs in five phases: 1. **Setup** — read the research profile and the persona template; produce a digest that confirms the right material is on the table. 2. **Creative ideation (the most important phase).** Before any section is filled, six creative tenets get locked through structured human-AI dialogue: - **Emotional core** — what feeling should come through in every recipe this person writes - **Era and setting** — contemporary, historical, or timeless - **What stays / transforms / gets invented** — every major biographical element is sorted into one of three buckets, deliberately, with the human's judgment driving the calls - **Narrative devices** — one or two invented anchors that give the voice something to return to (Chef Lupita's "mother's notebook" is one example) - **Voice direction** — the _feel_ of the prose, before any specific words are picked - **Heritage status** — whether this persona gets the deeper cultural and visual treatment 3. **Foundational decisions** — fictional name, archetype, output folder. 4. **Section-by-section translation.** Eleven persona sections are worked through in order. For each one, the agent extracts source material from the research, checks it against the locked tenets, identifies gaps requiring invention, presents two or three drafted options with trade-offs, and only writes when the human approves. The hardest sections (voice, few-shot examples, visual philosophy) take additional research passes — the visual philosophy section, for example, runs new web research on the cuisine's ceramics, plating traditions, and color palettes before any draft is written. 5. **Assembly and validation.** The persona is checked against quality gates: sentence burstiness, natural contractions, zero forbidden AI words (`delve`, `leverage`, `realm`, etc.), zero em dashes, signature phrases used naturally rather than forced. Seven full few-shot writing examples — recipe introductions, headlines, instruction passages, chef tips, anti-examples — serve as the voice proof-of-concept. For Austria, Gretel Beer's research became **Chef Elsa Wagner**: Austrian, Mehlspeisen specialist, rooted in Salzburg and Vienna. Her persona document is the contract — a multi-page specification of identity, philosophy, signature vocabulary, forbidden language specific to her voice, behavioral rules for common scenarios, visual philosophy, and seven worked examples. Every recipe she writes is generated against that specification. --- ## Stage 5 — Collection Plan Before Chef Elsa writes recipes, she plans her full body of work. The collection plan is a structured audit of all twelve recipe categories — appetizers, mains, desserts, breads, soups, and so on — written in her voice, deciding which categories belong in her tradition and which do not, and defining the themed collections within each. Every plan must define an **identity test** — a single question every dish must answer to be included. Chef Elsa's is "Why is this Austrian?" The test is the quality gate that makes filler impossible. The plan includes three explicit human approval points (organizing principle, collection list, identity test) where the human can redirect the chef before any recipes are committed to. --- ## Stage 6 — Recipe Collection (the two-pass research workflow) For each collection, the chef runs a two-pass curation workflow. Pass 2 exists because single-wave research returns "what comes to mind first," which is not the same as "what genuinely represents this category." Pass 2 is the structural mechanism that forces the chef back into the research before the file is written. **Pass 1 — Initial research.** Iterative search waves through the multi-provider tool: a broad landscape wave to map the category, a targeted wave on dishes the first wave surfaced, a gap-filling wave on what's still thin. The chef returns a first-cut dish list with cultural justifications and subcategory assignments. **Pass 2 — Second-look (non-negotiable, five sub-steps):** 1. **Restate the brief** — the chef articulates, in their own voice, what a complete representation of this category looks like in their tradition 2. **Reflect on the first cut** — explicitly identify which sub-areas are thin, what regional variations were skipped, what an outside expert who knows this tradition would say is missing 3. **Run gap-finding searches** aimed specifically at what the reflection surfaced (e.g., "What essential strudels in Austrian tradition are commonly overlooked outside Vienna?") 4. **Decide explicitly on every surfaced candidate** — each new dish is added with justification, or rejected with stated reason. Silent omission is forbidden. 5. **Decide keep-or-remove for every subcategory** — every subcategory listed in the plan is either kept (with at least one dish assigned) or removed (with a stated cultural-fit reason). "Already have enough" is not a valid reason. **Deterministic validation gates.** Before the file is written, every dish is checked against controlled vocabularies (categories, subcategories, cuisines, occasions). Dish names and SEO contexts are length-checked. Cross-collection duplicate checks run via grep against every other collection in the chef's folder. Failures are self-corrected; only irrecoverable failures escalate to a human. **Audit sidecar.** Every collection writes a JSON sidecar capturing every search wave, every Pass 2 candidate decision (added or rejected, with reason), every subcategory keep-or-remove decision, and the final validation outcome. The work can be reviewed, spot-checked, or audited after the fact. Nothing is invisible. For Chef Elsa's "Austrian Strudels" collection, Pass 2 surfaced regional variations and home-cook strudels that would never appear on a tourist menu but are central to the actual Austrian repertoire — and the sidecar records exactly which ones were added, which were rejected, and why. --- ## Stage 7 — The Published Recipe The recipe is generated against the chef's persona specification: their identity statement, signature vocabulary, forbidden language, behavioral rules, rhetorical patterns, tone calibration, and visual philosophy. The few-shot examples in the persona document anchor the voice. The dish itself comes from the collection. The cuisine, occasion, and category taxonomy come from the collection's validated metadata. The image is rendered against the chef's documented visual philosophy. By the time the recipe is written, every load-bearing decision has already been made and recorded upstream. The recipe-writing step itself is the smallest step in the pipeline. Sample output: [Apfelstrudel](https://culinaryexplorer.app/s/MgJLX-Hxoaat). --- ## Summary of artifacts Each stage produces durable, on-disk artifacts that constitute the canonical record of the work: | Stage | Primary Artifact | | ----------------- | ---------------------------------------------------------------------------- | | Global Council | Regional bloc map (YAML) + per-region council briefs + round transcripts | | Regional Council | Cuisine taxonomy (YAML) + selected authority per cuisine + round transcripts | | Chef Research | Sixteen-section research document + per-researcher findings audit folder | | Chef Persona | Persona specification (markdown) + manifest + installed prompt files | | Collection Plan | Strategy document (YAML) covering all twelve recipe categories | | Recipe Collection | Validated dish list (YAML) + run sidecar (JSON) capturing every decision | | Recipe | Published recipe page | Source-of-truth locations: - `C:\dev\culinary-advisor-admin\docs\plans\global-cuisine-council\` — global council deliberations - `C:\dev\culinary-advisor-admin\docs\plans\european-cuisine-council\` — regional council deliberations - `C:\dev\culinary-advisor-admin\docs\chefs\origins\` — research profiles for selected authorities - `C:\dev\culinary-advisor-admin\docs\chefs\personas\` — persona specifications, manifests, prompt files - `C:\dev\culinary-advisor-admin\docs\taxonomy\collections\` — collection plans, collection YAMLs, run sidecars - `C:\dev\culinary-advisor-admin\backend\agents\vocabularies\` — controlled vocabularies validated against - `C:\dev\culinary-advisor-admin\.claude\commands\` — workflow command definitions (research-chef, create-persona, install-chef, new-collection-plan, new-collection-v2)