refactor(ai): collapse parser prompts to template; replace makeEmptyItem with structuredClone

- Replace pdf-parser-system.md + docx-parser-system.md with a single
  parser-system.md template; prompts.ts substitutes 6 placeholders per
  source type. Produced strings are byte-identical to the former files.
- Remove makeEmptyItem recursive walker (all SECTION_ITEM_SHAPES leaves are
  already zero-valued) and call structuredClone(shape) instead.

Claude-Session: https://claude.ai/code/session_012Bnvt1MghwHj4qQRxuQUGa

packages/ai/src/prompts.ts

@@ -4,11 +4,46 @@ const readPrompt = (filename: string) => {
 	return readFileSync(new URL(`./prompts/${filename}`, import.meta.url), "utf-8");
 };
 
+// ponytail: single template with per-source substitutions; produced text is identical per source type
+const parserTemplate = readPrompt("parser-system.md");
+
+type ParserVars = {
+	FORMAT_HEADER: string;
+	FORMAT_NOUN: string;
+	ALLOWED_INPUT: string;
+	URL_CLAUSE: string;
+	EXTRA_RULES: string;
+	FALLBACK_CLAUSE: string;
+};
+
+function makeParserPrompt(vars: ParserVars): string {
+	return Object.entries(vars).reduce((acc, [k, v]) => acc.replaceAll(`{{${k}}}`, v), parserTemplate);
+}
+
+const pdfParserSystemPrompt = makeParserPrompt({
+	FORMAT_HEADER: "PDF files",
+	FORMAT_NOUN: "PDF",
+	ALLOWED_INPUT:
+		"- Use only the visible content from the attached PDF document.\n- Ignore OCR noise, watermarks, repeated headers/footers, and broken line wraps.",
+	URL_CLAUSE: "full URLs that are explicitly present",
+	EXTRA_RULES: "",
+	FALLBACK_CLAUSE: "PDF is low quality or partially unreadable",
+});
+
+const docxParserSystemPrompt = makeParserPrompt({
+	FORMAT_HEADER: "Microsoft Word files (DOC/DOCX)",
+	FORMAT_NOUN: "document",
+	ALLOWED_INPUT:
+		"- Use only visible, intended content from the attached document.\n- Ignore hidden text, comments, track changes, revision history, document metadata, and layout artifacts.",
+	URL_CLAUSE: "URLs explicitly visible in document content",
+	EXTRA_RULES:
+		"- Lists and tables: extract visible text faithfully; preserve relationships in section fields.\n- Headers/footers: include only if they contain real resume data.\n",
+	FALLBACK_CLAUSE: "document is malformed or partially unreadable",
+});
+
 const analyzeResumeSystemPrompt = readPrompt("analyze-resume-system.md");
 const chatSystemPromptTemplate = readPrompt("chat-system.md");
-const docxParserSystemPrompt = readPrompt("docx-parser-system.md");
 const docxParserUserPrompt = readPrompt("docx-parser-user.md");
-const pdfParserSystemPrompt = readPrompt("pdf-parser-system.md");
 const pdfParserUserPrompt = readPrompt("pdf-parser-user.md");
 
 export {

packages/ai/src/prompts/docx-parser-system.md

@@ -1,54 +0,0 @@
-You are a strict resume extraction engine for Microsoft Word files (DOC/DOCX). Convert the attached document into a Reactive Resume JSON object.
-
-## Objective
-
-- Extract resume content accurately and map it into the provided JSON template.
-- Prioritize source fidelity and schema correctness over completeness.
-
-## Allowed Input
-
-- Use only visible, intended content from the attached document.
-- Ignore hidden text, comments, track changes, revision history, document metadata, and layout artifacts.
-
-## Hard Constraints
-
-1. Extract only explicitly stated information.
-2. Never fabricate, infer, or normalize missing data.
-3. Keep original wording and original language.
-4. When uncertain, omit content and leave template defaults.
-5. Do not use external knowledge.
-
-## Conflict Resolution Order
-
-1. Schema validity (must return valid JSON matching template shape)
-2. Source fidelity (exactly what the document states)
-3. Omit uncertain values (never guess)
-
-## Extraction Rules
-
-- Dates: preserve exactly as written.
-- URLs: include only URLs explicitly visible in document content.
-- Contact data: copy as-is; do not reformat.
-- Skills: include only explicit skill mentions.
-- Descriptions: output HTML using `<p>`, `<ul>`, `<li>` while preserving meaning.
-- Lists and tables: extract visible text faithfully; preserve relationships in section fields.
-- Headers/footers: include only if they contain real resume data.
-- IDs: generate unique UUIDs for all `id` fields.
-- `hidden`: default to `false` unless explicitly indicated otherwise.
-- `columns`: default to `1` unless clearly multi-column by content intent.
-- `website`: when missing, use `{ "url": "", "label": "" }`.
-
-## Section Mapping
-
-- `basics`, `summary`, `experience`, `education`, `skills`, `projects`, `certifications`, `awards`, `languages`, `volunteer`, `publications`, `references`, `profiles`, `interests`
-- Map based on explicit headings first; use local context only when heading is absent.
-
-## Fallback Rules
-
-- If the document is malformed or partially unreadable, return best-effort extraction for readable parts only.
-- Keep unknown fields empty according to the template.
-
-## Output Contract
-
-- Return only one raw JSON object.
-- No markdown, no commentary, no extra keys.

packages/ai/src/prompts/parser-system.md

@@ -1,4 +1,4 @@
-You are a strict resume extraction engine for PDF files. Convert the attached PDF into a Reactive Resume JSON object.
+You are a strict resume extraction engine for {{FORMAT_HEADER}}. Convert the attached {{FORMAT_NOUN}} into a Reactive Resume JSON object.
 
 ## Objective
 
@@ -7,8 +7,7 @@ You are a strict resume extraction engine for PDF files. Convert the attached PD
 
 ## Allowed Input
 
-- Use only the visible content from the attached PDF document.
-- Ignore OCR noise, watermarks, repeated headers/footers, and broken line wraps.
+{{ALLOWED_INPUT}}
 
 ## Hard Constraints
 
@@ -21,17 +20,17 @@ You are a strict resume extraction engine for PDF files. Convert the attached PD
 ## Conflict Resolution Order
 
 1. Schema validity (must return valid JSON matching template shape)
-2. Source fidelity (exactly what the PDF states)
+2. Source fidelity (exactly what the {{FORMAT_NOUN}} states)
 3. Omit uncertain values (never guess)
 
 ## Extraction Rules
 
 - Dates: preserve exactly as written.
-- URLs: include only full URLs that are explicitly present.
+- URLs: include only {{URL_CLAUSE}}.
 - Contact data: copy as-is; do not reformat.
 - Skills: include only explicit skill mentions.
 - Descriptions: output HTML using `<p>`, `<ul>`, `<li>` while preserving meaning.
-- IDs: generate unique UUIDs for all `id` fields.
+{{EXTRA_RULES}}- IDs: generate unique UUIDs for all `id` fields.
 - `hidden`: default to `false` unless explicitly indicated otherwise.
 - `columns`: default to `1` unless clearly multi-column by content intent.
 - `website`: when missing, use `{ "url": "", "label": "" }`.
@@ -43,7 +42,7 @@ You are a strict resume extraction engine for PDF files. Convert the attached PD
 
 ## Fallback Rules
 
-- If the PDF is low quality or partially unreadable, return best-effort extraction for readable parts only.
+- If the {{FALLBACK_CLAUSE}}, return best-effort extraction for readable parts only.
 - Keep unknown fields empty according to the template.
 
 ## Output Contract

packages/ai/src/resume/extraction-template.ts

@@ -1,23 +1,5 @@
 import { defaultResumeData } from "@reactive-resume/schema/resume/default";
 
-function makeEmptyItem(shape: Record<string, unknown>): Record<string, unknown> {
-	const item: Record<string, unknown> = {};
-
-	for (const [key, value] of Object.entries(shape)) {
-		if (typeof value === "string") item[key] = "";
-		else if (typeof value === "number") item[key] = 0;
-		else if (typeof value === "boolean") item[key] = false;
-		else if (Array.isArray(value)) item[key] = [];
-		else if (value !== null && typeof value === "object") {
-			item[key] = makeEmptyItem(value as Record<string, unknown>);
-		} else {
-			item[key] = "";
-		}
-	}
-
-	return item;
-}
-
 const SECTION_ITEM_SHAPES: Record<string, Record<string, unknown>> = {
 	profiles: {
 		id: "",
@@ -114,7 +96,7 @@ export function buildAiExtractionTemplate() {
 		const sectionKey = key as keyof typeof defaultResumeData.sections;
 		sections[key] = {
 			...defaultResumeData.sections[sectionKey],
-			items: [makeEmptyItem(shape)],
+			items: [structuredClone(shape)],
 		};
 	}