extract-synthesis-recipe

name	extract-synthesis-recipe
description	Extracts materials-synthesis recipes from free-form text (papers, abstracts, lab notes) into a strict JSON schema. Use when the user asks to parse, extract, or structure synthesis information, or wants to convert experimental procedures into machine-readable form.
when_to_use	Synthesis extraction tasks, parsing methods sections, building synthesis databases, populating reconstruction or simulation configs from prose.
allowed-tools	Read

Extract Synthesis Recipe

Task

Read the text the user provides (or a file they point to) and extract every distinct synthesis procedure as a JSON record with the schema below. If the text contains multiple recipes, return a JSON array; otherwise return a single object.

Output schema

{
  "material":            "string  (chemical formula)",
  "route":               "sol-gel | solid-state | hydrothermal | co-precipitation | CVD | other",
  "precursors":          ["array of strings"],
  "calcination_temp_C":  "number | null",
  "calcination_time_h":  "number | null",
  "atmosphere":          "string | null",
  "crystallite_size_nm": "number | null",
  "confidence":          "high | medium | low",
  "evidence_quote":      "verbatim quote from the source supporting the extraction",
  "extraction_notes":    "string | null  (use to flag ambiguities or missing fields)"
}

Rules

1. Never invent values. If a field is not stated in the text, set it to null and lower confidence.
2. Quote the evidence. evidence_quote must be a verbatim substring of the source — do not paraphrase.
3. Flag non-syntheses. If the passage is not actually a synthesis (e.g. a modelling paper, an instrumentation description), do not invent a record. Instead, return a single object with confidence: "low" and extraction_notes explaining why.
4. Preserve units. Convert "overnight" → null + a note; do not guess "8 h".
5. Chemistry is verbatim. Keep formula casing/subscripts as written (BaTiO3 not batio3).

Validation checklist (run mentally before returning)

• Did I quote evidence for every non-null field?
• Did I refuse to fabricate values for missing fields?
• Did I flag any decoy passages with confidence: low?
• Is the JSON parseable (no trailing commas, double quotes only)?

Output

Return only the JSON. No commentary, no markdown fences. The user's pipeline will parse your response directly.