recipe research

recipe-research

"Find me a recipe" gets you one blogger's opinion. A restaurant-level recipe needs three things a single search never gives you together: what several well-regarded sources actually do (and where they disagree), the food science behind why the good version works, and what professional kitchens do differently from home cooks.

This plugin runs all three as parallel research, then makes the editorial call — which disagreement resolves which way, and why — instead of averaging opinions into mush.

Setup

/plugin marketplace add roypadina/padina-claude-code-plugins
/plugin install recipe-research@padina

Requires nothing but Claude Code. No CLI, no runtime, no account. It writes Markdown files and that is all. Obsidian is optional and off by default.

What you get

Feature	Mechanism
/recipe-research <dish>	slash command → the skill
Three parallel research subagents	dispatched at runtime by the skill
A synthesized recipe in a linked note tree	written to <outputDir>/<cuisine>/<dish>/

No hooks, no bundled agent files, no MCP server — the three researchers are ordinary general-purpose subagents the skill fans out in one dispatch, with the model tier picked per agent.

Usage

/recipe-research shakshuka

Or just ask: "research a recipe for hummus", "find me a restaurant-level carbonara", "add shakshuka to my recipes".

The three researchers

Agent	Brief
classic recipes	What several well-regarded sources actually do — attributed, quantities exact, disagreements recorded rather than smoothed over
food science	Why the good version works, and the failure modes of the bad one
pro-chef technique	Named chefs and restaurant practice — anything that cannot be pinned to a named source is flagged UNVERIFIED rather than asserted

They run in one dispatch, in parallel. The main thread then does the synthesis: golden rules, metric quantities, exact temps and times, and a "why this way" section that justifies each call against the research.

On "sourced". The subagents work from model knowledge, and search the web when they have those tools — the skill does not require a search tool. Read "sourced" as attributed, not as guaranteed-live-fetched.

Output structure

<outputDir>/<cuisine>/<dish>/
  <dish>.md                     # synthesized recipe — the hub
  research/
    classic-recipes.md          # summary: recipes compared
    food-science.md             # summary: why it works
    pro-technique.md            # summary: pro/restaurant technique
    raw/
      classic-recipes-raw.md    # verbatim subagent output
      food-science-raw.md
      pro-technique-raw.md

The recipe links to the three summaries; each summary links to its own raw file. Nothing is thrown away — you can always read what an agent actually said before it was condensed.

Configuration

/plugin configure recipe-research

Option	Type	Default
outputDir	directory	~/recipes	Where dish folders are written
vaultMode	boolean	false	true for Obsidian [[wiki-links]] + frontmatter tags; false for plain relative Markdown links
language	string	auto	Language for the recipe and summaries; auto matches whatever language you are writing in

Or set them at install:

/plugin install recipe-research@padina --config outputDir=~/recipes --config vaultMode=true

Obsidian — optional, off by default

Nothing checks for an Obsidian install, and none is needed. Plain mode writes relative Markdown links that read correctly in any editor or on GitHub.

If you do keep a vault, point outputDir at a folder inside it and set vaultMode=true. The same note tree is then written with [[wiki-links]] and tags / source-agent / date frontmatter, so Obsidian's graph view and tag search pick it up natively. Nothing else about the workflow changes.

Uninstalling

/plugin uninstall recipe-research@padina

Your recipe folder stays where it is — the plugin only ever wrote into outputDir.