Agentic Recipes
Not to be confused with Drupal core recipes (declarative
recipe.ymlconfig packages). An agentic recipe is the AI-native equivalent: a set of agentic steps carrying the reasoning, derivation, gating, and invocation a declarative format cannot. One recipe delivers one named capability end to end.
A recipe is prescriptive, not descriptive. "How X works in Drupal" is a guide; "how we do X, and what we refuse to do" is a recipe. A recipe cites guides and plays, never duplicates them, and is not a recipe at all unless it carries a name an agent can route to, the constraint it enforces, and a verifier that catches the agent when it drifts.
Recipes are published to a separate index, agentic-recipes.txt (not llms.txt), exposing only each recipe's routing block (name / capability / description). A caller matches a capability first, loads the recipe, and the recipe names the guides it needs.
Routing table
| Capability | Recipe | When to use |
|---|---|---|
responsive-image-delivery |
Responsive image wiring | A Drupal site has named image use-cases (hero, card thumbnail, content inline) that must render as responsive images on image fields. |
See also
- File-format standard:
recipe_schema_version 1.0.0— routing-first frontmatter, fixed body section set, verifier required. - Machine-readable dependencies (optional): a recipe body may declare
requires_guides:andrequires_plays:— lists of guide/play slugs (e.g.drupal/image-styles/image-overview) the recipe-loader resolves without parsing prose. Each slug must resolve to a published guide/play (docs/<slug>.mdordocs/<slug>/index.md);scripts/validate_recipes.py(run in CI) fails the build on a dangling slug. A recipe that omits them still matches — its aspects fall through to residual guide-search. These keys live in the recipe body (fetched raw on match), not the routing index. - Drupal best-practice plays — the prescriptive stances recipes cite as sources.