Recipe Remix Primer
Recipe remixing is the practice of taking a working prompt chain, model route, or agent recipe and shaping it into a new flow without rebuilding from zero. This primer walks through the three remix patterns Meridian users reach for most often, and the tradeoffs each one carries when you wire them into a production pipeline.
01.Fork and rewire
The simplest remix is a fork. You clone an existing recipe, then swap one node at a time and re-run the eval suite after each swap. This keeps the blast radius small and lets you bisect regressions cleanly. Forking shines when you trust the structure but want a different model or a tighter system prompt.
Use forking when the recipe topology is sound and you only need to tune a step or two.
02.Splice and chain
Splicing takes two recipes and joins them at a shared schema. The output of recipe A becomes the input of recipe B without translation. Meridian enforces a contract check at the splice point, so a schema drift fails fast instead of silently passing malformed data downstream.
// splice two recipes at a shared schema
const remix = compose(
recipes.extractEntities, // -> { entities: Entity[] }
recipes.summarizeEntities // expects { entities: Entity[] }
);
const result = await remix.run(input);03.Parameterize and reuse
The third pattern lifts hardcoded values into recipe parameters. A summarizer recipe with a baked-in word count becomes a summarizer recipe that takes a word count as input. Once a recipe is parameterized, the same artifact serves a dozen callers and the eval cost amortizes across all of them.
Parameterization is the highest-leverage remix because it converts a one-off into a reusable primitive.