Developer Role AI-Native Shift-Left Playbook

Elevating prompt development into a continuous, version-controlled, CI-gated engineering lifecycle.

Delineating probabilistic model reasoning from developer-owned deterministic harness enforcement.

// ❌ ANTI-PATTERN: Inline string in business logic, zero typing, brittle parsing
export async function summarizePatientRecord(noteText: string) {
  // Unversioned, untracked prompt string
  const prompt = "Please read this doctor note: " + noteText + 
    " and tell me the diagnosis, key symptoms, and if it is urgent. Format as JSON.";
  
  const response = await openai.chat.completions.create({
    model: "gpt-4o",
    messages: [{ role: "user", content: prompt }]
  });

  const rawText = response.choices[0].message.content || "{}";
  
  // Brittle assumption that model returns clean JSON without markdown fences
  try {
    return JSON.parse(rawText);
  } catch (err) {
    // Silent failure or raw crash in production
    console.error("JSON parse error", err);
    return { error: "Failed to parse" };
  }
}

• Unversioned Logic: Modifying the prompt string leaves no git history of prompt performance or regression benchmarks.
• Missing Context Delimiters: Physician note text is directly concatenated without XML tag isolation, exposing the system to prompt injection.
• Zero Schema Enforcement: Returns raw, unvalidated JSON that frequently crashes frontend components when keys are omitted or misnamed.
• No Automated CI Testing: Tested only manually in a web browser; zero automated checks against edge cases or complex medical conditions.

// ✅ SHIFT-LEFT: Modular Prompt-as-Code asset + strict Zod deterministic harness
import { z } from 'zod';
import { CLINICAL_SUMMARY_PROMPT_V2 } from '@/prompts/clinicalSummaryPrompt.v2';
import { callModelWithTelemetry } from '@/lib/ai/modelClient';

export const ClinicalSummarySchema = z.object({
  patientId: z.string().uuid(),
  icd10Codes: z.array(z.string().regex(/^[A-Z][0-9]{2}(\.[0-9]{1,2})?$/)),
  primaryDiagnosis: z.string().min(2),
  keySymptoms: z.array(z.string()).min(1),
  urgencyLevel: z.enum(['ROUTINE', 'URGENT', 'CRITICAL']),
  confidenceScore: z.number().min(0).max(1),
});

export type ClinicalSummary = z.infer<typeof ClinicalSummarySchema>;

export async function extractClinicalSummary(
  patientId: string, 
  physicianNote: string
): Promise<ClinicalSummary> {
  // 1. Build typed prompt from versioned asset
  const formattedPrompt = CLINICAL_SUMMARY_PROMPT_V2.format({
    patientId,
    sanitizedNote: sanitizeInput(physicianNote),
  });

  // 2. Call model with structured JSON mode and telemetry
  const rawOutput = await callModelWithTelemetry({
    prompt: formattedPrompt,
    promptVersion: CLINICAL_SUMMARY_PROMPT_V2.version, // e.g. "2.1.0"
    temperature: 0.1,
  });

  // 3. Strict Deterministic Schema Validation
  const parseResult = ClinicalSummarySchema.safeParse(rawOutput);
  if (parseResult.success) {
    return parseResult.data;
  }

  // 4. Deterministic self-correcting retry or safe fallback
  return handleSchemaCorrection(formattedPrompt, parseResult.error, patientId);
}

• Modular Versioned Asset: Prompt lives in `src/prompts/` with semantic versioning (`v2.1.0`) and full git traceability.
• Strict Type & Schema Safety: Zod schema enforces exact regex patterns, required arrays, and non-empty strings before passing data to downstream code.
• Injection & Delimiter Protection: Input notes are sanitized and wrapped in structured XML delimiters (`<physician_note>`), preventing prompt hijacking.
• CI Regression Gated: All modifications to `clinicalSummaryPrompt.v2.ts` automatically trigger the 50-case golden-set CI evaluation suite before merge.

Ready to Benchmark Your Team’s Prompt Discipline?